[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"similar-Meirtz--Awesome-Context-Engineering":3,"tool-Meirtz--Awesome-Context-Engineering":62},[4,18,26,36,46,54],{"id":5,"name":6,"github_repo":7,"description_zh":8,"stars":9,"difficulty_score":10,"last_commit_at":11,"category_tags":12,"status":17},4358,"openclaw","openclaw\u002Fopenclaw","OpenClaw 是一款专为个人打造的本地化 AI 助手，旨在让你在自己的设备上拥有完全可控的智能伙伴。它打破了传统 AI 助手局限于特定网页或应用的束缚，能够直接接入你日常使用的各类通讯渠道，包括微信、WhatsApp、Telegram、Discord、iMessage 等数十种平台。无论你在哪个聊天软件中发送消息，OpenClaw 都能即时响应，甚至支持在 macOS、iOS 和 Android 设备上进行语音交互，并提供实时的画布渲染功能供你操控。\n\n这款工具主要解决了用户对数据隐私、响应速度以及“始终在线”体验的需求。通过将 AI 部署在本地，用户无需依赖云端服务即可享受快速、私密的智能辅助，真正实现了“你的数据，你做主”。其独特的技术亮点在于强大的网关架构，将控制平面与核心助手分离，确保跨平台通信的流畅性与扩展性。\n\nOpenClaw 非常适合希望构建个性化工作流的技术爱好者、开发者，以及注重隐私保护且不愿被单一生态绑定的普通用户。只要具备基础的终端操作能力（支持 macOS、Linux 及 Windows WSL2），即可通过简单的命令行引导完成部署。如果你渴望拥有一个懂你",349277,3,"2026-04-06T06:32:30",[13,14,15,16],"Agent","开发框架","图像","数据工具","ready",{"id":19,"name":20,"github_repo":21,"description_zh":22,"stars":23,"difficulty_score":10,"last_commit_at":24,"category_tags":25,"status":17},3808,"stable-diffusion-webui","AUTOMATIC1111\u002Fstable-diffusion-webui","stable-diffusion-webui 是一个基于 Gradio 构建的网页版操作界面，旨在让用户能够轻松地在本地运行和使用强大的 Stable Diffusion 图像生成模型。它解决了原始模型依赖命令行、操作门槛高且功能分散的痛点，将复杂的 AI 绘图流程整合进一个直观易用的图形化平台。\n\n无论是希望快速上手的普通创作者、需要精细控制画面细节的设计师，还是想要深入探索模型潜力的开发者与研究人员，都能从中获益。其核心亮点在于极高的功能丰富度：不仅支持文生图、图生图、局部重绘（Inpainting）和外绘（Outpainting）等基础模式，还独创了注意力机制调整、提示词矩阵、负向提示词以及“高清修复”等高级功能。此外，它内置了 GFPGAN 和 CodeFormer 等人脸修复工具，支持多种神经网络放大算法，并允许用户通过插件系统无限扩展能力。即使是显存有限的设备，stable-diffusion-webui 也提供了相应的优化选项，让高质量的 AI 艺术创作变得触手可及。",162132,"2026-04-05T11:01:52",[14,15,13],{"id":27,"name":28,"github_repo":29,"description_zh":30,"stars":31,"difficulty_score":32,"last_commit_at":33,"category_tags":34,"status":17},1381,"everything-claude-code","affaan-m\u002Feverything-claude-code","everything-claude-code 是一套专为 AI 编程助手（如 Claude Code、Codex、Cursor 等）打造的高性能优化系统。它不仅仅是一组配置文件，而是一个经过长期实战打磨的完整框架，旨在解决 AI 代理在实际开发中面临的效率低下、记忆丢失、安全隐患及缺乏持续学习能力等核心痛点。\n\n通过引入技能模块化、直觉增强、记忆持久化机制以及内置的安全扫描功能，everything-claude-code 能显著提升 AI 在复杂任务中的表现，帮助开发者构建更稳定、更智能的生产级 AI 代理。其独特的“研究优先”开发理念和针对 Token 消耗的优化策略，使得模型响应更快、成本更低，同时有效防御潜在的攻击向量。\n\n这套工具特别适合软件开发者、AI 研究人员以及希望深度定制 AI 工作流的技术团队使用。无论您是在构建大型代码库，还是需要 AI 协助进行安全审计与自动化测试，everything-claude-code 都能提供强大的底层支持。作为一个曾荣获 Anthropic 黑客大奖的开源项目，它融合了多语言支持与丰富的实战钩子（hooks），让 AI 真正成长为懂上",160015,2,"2026-04-18T11:30:52",[14,13,35],"语言模型",{"id":37,"name":38,"github_repo":39,"description_zh":40,"stars":41,"difficulty_score":42,"last_commit_at":43,"category_tags":44,"status":17},8272,"opencode","anomalyco\u002Fopencode","OpenCode 是一款开源的 AI 编程助手（Coding Agent），旨在像一位智能搭档一样融入您的开发流程。它不仅仅是一个代码补全插件，而是一个能够理解项目上下文、自主规划任务并执行复杂编码操作的智能体。无论是生成全新功能、重构现有代码，还是排查难以定位的 Bug，OpenCode 都能通过自然语言交互高效完成，显著减少开发者在重复性劳动和上下文切换上的时间消耗。\n\n这款工具专为软件开发者、工程师及技术研究人员设计，特别适合希望利用大模型能力来提升编码效率、加速原型开发或处理遗留代码维护的专业人群。其核心亮点在于完全开源的架构，这意味着用户可以审查代码逻辑、自定义行为策略，甚至私有化部署以保障数据安全，彻底打破了传统闭源 AI 助手的“黑盒”限制。\n\n在技术体验上，OpenCode 提供了灵活的终端界面（Terminal UI）和正在测试中的桌面应用程序，支持 macOS、Windows 及 Linux 全平台。它兼容多种包管理工具，安装便捷，并能无缝集成到现有的开发环境中。无论您是追求极致控制权的资深极客，还是渴望提升产出的独立开发者，OpenCode 都提供了一个透明、可信",144296,1,"2026-04-16T14:50:03",[13,45],"插件",{"id":47,"name":48,"github_repo":49,"description_zh":50,"stars":51,"difficulty_score":32,"last_commit_at":52,"category_tags":53,"status":17},2271,"ComfyUI","Comfy-Org\u002FComfyUI","ComfyUI 是一款功能强大且高度模块化的视觉 AI 引擎，专为设计和执行复杂的 Stable Diffusion 图像生成流程而打造。它摒弃了传统的代码编写模式，采用直观的节点式流程图界面，让用户通过连接不同的功能模块即可构建个性化的生成管线。\n\n这一设计巧妙解决了高级 AI 绘图工作流配置复杂、灵活性不足的痛点。用户无需具备编程背景，也能自由组合模型、调整参数并实时预览效果，轻松实现从基础文生图到多步骤高清修复等各类复杂任务。ComfyUI 拥有极佳的兼容性，不仅支持 Windows、macOS 和 Linux 全平台，还广泛适配 NVIDIA、AMD、Intel 及苹果 Silicon 等多种硬件架构，并率先支持 SDXL、Flux、SD3 等前沿模型。\n\n无论是希望深入探索算法潜力的研究人员和开发者，还是追求极致创作自由度的设计师与资深 AI 绘画爱好者，ComfyUI 都能提供强大的支持。其独特的模块化架构允许社区不断扩展新功能，使其成为当前最灵活、生态最丰富的开源扩散模型工具之一，帮助用户将创意高效转化为现实。",109154,"2026-04-18T11:18:24",[14,15,13],{"id":55,"name":56,"github_repo":57,"description_zh":58,"stars":59,"difficulty_score":32,"last_commit_at":60,"category_tags":61,"status":17},6121,"gemini-cli","google-gemini\u002Fgemini-cli","gemini-cli 是一款由谷歌推出的开源 AI 命令行工具，它将强大的 Gemini 大模型能力直接集成到用户的终端环境中。对于习惯在命令行工作的开发者而言，它提供了一条从输入提示词到获取模型响应的最短路径，无需切换窗口即可享受智能辅助。\n\n这款工具主要解决了开发过程中频繁上下文切换的痛点，让用户能在熟悉的终端界面内直接完成代码理解、生成、调试以及自动化运维任务。无论是查询大型代码库、根据草图生成应用，还是执行复杂的 Git 操作，gemini-cli 都能通过自然语言指令高效处理。\n\n它特别适合广大软件工程师、DevOps 人员及技术研究人员使用。其核心亮点包括支持高达 100 万 token 的超长上下文窗口，具备出色的逻辑推理能力；内置 Google 搜索、文件操作及 Shell 命令执行等实用工具；更独特的是，它支持 MCP（模型上下文协议），允许用户灵活扩展自定义集成，连接如图像生成等外部能力。此外，个人谷歌账号即可享受免费的额度支持，且项目基于 Apache 2.0 协议完全开源，是提升终端工作效率的理想助手。",100752,"2026-04-10T01:20:03",[45,13,15,14],{"id":63,"github_repo":64,"name":65,"description_en":66,"description_zh":67,"ai_summary_zh":68,"readme_en":69,"readme_zh":70,"quickstart_zh":71,"use_case_zh":72,"hero_image_url":73,"owner_login":74,"owner_name":75,"owner_avatar_url":76,"owner_bio":77,"owner_company":78,"owner_location":79,"owner_email":80,"owner_twitter":81,"owner_website":82,"owner_url":83,"languages":81,"stars":84,"forks":85,"last_commit_at":86,"license":87,"difficulty_score":42,"env_os":88,"env_gpu":89,"env_ram":89,"env_deps":90,"category_tags":93,"github_topics":95,"view_count":32,"oss_zip_url":81,"oss_zip_packed_at":81,"status":17,"created_at":104,"updated_at":105,"faqs":106,"releases":107},9299,"Meirtz\u002FAwesome-Context-Engineering","Awesome-Context-Engineering"," 🔥 Comprehensive survey on Context Engineering: from prompt engineering to production-grade AI systems. hundreds of papers, frameworks, and  implementation guides for LLMs and AI agents.","Awesome-Context-Engineering 是一个专注于“上下文工程”的开源资源库，旨在帮助开发者构建从基础提示词优化到生产级 AI 系统的完整能力。随着大语言模型（LLM）应用的深入，传统的静态提示词已难以应对复杂任务的不确定性，而 Awesome-Context-Engineering 正是为了解决这一痛点而生。它系统性地整理了数百篇学术论文、主流框架及实战指南，涵盖了动态上下文管理、检索增强生成（RAG）、记忆系统、智能体（Agent）运行时、工具调用协议以及可观测性栈等关键技术领域。\n\n该项目不仅适合希望提升模型输出质量的应用开发者，也深受 AI 研究人员和架构师的青睐。对于正在探索如何让人工智能代理具备长期规划、状态管理和人机协作能力的团队，这里提供了从理论综述到代码落地的全方位参考。其独特亮点在于紧跟技术前沿，特别更新了面向 2026 年“智能体时代”的内容，深入探讨了智能体编排、持久化记忆工件及生产环境下的上下文压缩策略。无论你是想入门上下文工程，还是寻求构建高可靠性 AI 应用的深度指导，Awesome-Context-Engineering 都是一份不可多","Awesome-Context-Engineering 是一个专注于“上下文工程”的开源资源库，旨在帮助开发者构建从基础提示词优化到生产级 AI 系统的完整能力。随着大语言模型（LLM）应用的深入，传统的静态提示词已难以应对复杂任务的不确定性，而 Awesome-Context-Engineering 正是为了解决这一痛点而生。它系统性地整理了数百篇学术论文、主流框架及实战指南，涵盖了动态上下文管理、检索增强生成（RAG）、记忆系统、智能体（Agent）运行时、工具调用协议以及可观测性栈等关键技术领域。\n\n该项目不仅适合希望提升模型输出质量的应用开发者，也深受 AI 研究人员和架构师的青睐。对于正在探索如何让人工智能代理具备长期规划、状态管理和人机协作能力的团队，这里提供了从理论综述到代码落地的全方位参考。其独特亮点在于紧跟技术前沿，特别更新了面向 2026 年“智能体时代”的内容，深入探讨了智能体编排、持久化记忆工件及生产环境下的上下文压缩策略。无论你是想入门上下文工程，还是寻求构建高可靠性 AI 应用的深度指导，Awesome-Context-Engineering 都是一份不可多得的权威地图。","# Awesome Context Engineering\n\n\u003Cdiv align=\"center\">\n  \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FMeirtz_Awesome-Context-Engineering_readme_f88855f12e97.png\" alt=\"Awesome Context Engineering Cover\" width=\"800\"\u002F>\n\u003C\u002Fdiv>\n\n## 💬 Join Our Community\n\n\u003Cdiv align=\"center\">\n  \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FMeirtz_Awesome-Context-Engineering_readme_378aa8ccb7ca.png\" alt=\"WeChat Group\" width=\"200\"\u002F>\n  \u003Cp>\u003Cstrong>Join our WeChat group for discussions and updates!\u003C\u002Fstrong>\u003C\u002Fp>\n  \u003Cp>\u003Ca href=\"https:\u002F\u002Fdiscord.gg\u002Ffsqs3Ybh\">\u003Cstrong>Join our Discord server\u003C\u002Fstrong>\u003C\u002Fa>\u003C\u002Fp>\n\u003C\u002Fdiv>\n\n[![Awesome](https:\u002F\u002Fawesome.re\u002Fbadge.svg)](https:\u002F\u002Fawesome.re)\n[![License: MIT](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLicense-MIT-yellow.svg)](https:\u002F\u002Fopensource.org\u002Flicenses\u002FMIT)\n[![PRs Welcome](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FPRs-welcome-brightgreen.svg?style=flat-square)](http:\u002F\u002Fmakeapullrequest.com)\n[![Paper](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FPaper-Published-green.svg)](https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.13334)\n\n> 📄 **Our comprehensive survey paper on Context Engineering is now published!** Check out our latest academic insights and theoretical foundations.\n\nA comprehensive survey and collection of resources on **Context Engineering** - the evolution from static prompting to dynamic, context-aware AI systems, and increasingly to **agent runtimes, memory systems, protocols, coding agents, and observability stacks**.\n\n## 📧 Contact\n\nFor questions, suggestions, or collaboration opportunities, please feel free to reach out:\n\n**Lingrui Mei**  \n📧 Email:  [meilingrui25b@ict.ac.cn](mailto:meilingrui25b@ict.ac.cn) or [meilingrui22@mails.ucas.ac.cn](mailto:meilingrui22@mails.ucas.ac.cn)\n\n**I WROTE THE WRONG EMAIL ADDRESS IN THE FIRST VERSION OF MY PAPER!!** You can also open an issue in this repository for general discussions and suggestions.\n\n---\n\n## 📰 News\n\n- **[2025.07.17]** 🔥🔥 Our paper is now published! Check out [\"A Survey of Context Engineering for Large Language Models\"](https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.13334) on [arXiv](https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.13334) and [Hugging Face Papers](https:\u002F\u002Fhuggingface.co\u002Fpapers\u002F2507.13334)\n- **[2025.07.03]** Repository initialized with comprehensive outline\n- **[2025.07.03]** Survey structure established following modern context engineering paradigms\n\n---\n\n## 🎯 Introduction\n\nIn the era of Large Language Models (LLMs), the limitations of static prompting have become increasingly apparent. **Context Engineering** represents the natural evolution to address LLM uncertainty and achieve production-grade AI deployment. Unlike traditional prompt engineering, context engineering encompasses the complete information payload provided to LLMs at inference time, including all structured informational components necessary for plausible task completion.\n\nThis repository serves as a comprehensive survey of context engineering techniques, methodologies, and applications.\n\n---\n\n## 🧭 2026 Agent Era Update\n\n### From Context Engineering to Agent Engineering\n\nAs of **March 2026**, context engineering remains a useful and necessary concept, but it is no longer the whole story. The center of gravity has shifted from \"how to pack the best prompt\" to **how agent systems manage runtime state, memory, tools, protocols, approvals, and long-horizon execution**. In practice, context engineering now sits inside a broader stack that also includes **agent harnesses**, **interoperability protocols**, **project memory for coding agents**, and **trace-first observability**.\n\n### What This Repository Now Covers\n\nThis repository still preserves its original survey structure on long context, RAG, memory, agent communication, tool use, evaluation, and applications. At the same time, this README is being reorganized to better reflect the **agent era** through additional coverage of:\n\n- **Agent harnesses and runtime systems** for planning, subagents, checkpoints, sandboxes, and human approval loops\n- **Context management in production** through compaction, caching, artifact-backed context, and scoped instruction loading\n- **Memory artifacts and portability** including persistent memory, memory interchange formats, persona packaging, and project memory\n- **Open protocols** such as MCP, A2A, AG-UI, ACP, and portable agent schemas\n- **Coding agents and computer use** as the most visible production setting for context engineering today\n- **Evaluation, observability, and telemetry** for long-running agent systems rather than only static benchmarks\n\n### Reading Guide for 2026 Topics\n\nReaders primarily interested in the 2026 shift should jump to the expanded sections on:\n\n- **Agent harnesses and runtime systems**, inspired by [Anthropic's effective agents guide](https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Fbuilding-effective-agents), [OpenAI's Agents and Tools documentation](https:\u002F\u002Fplatform.openai.com\u002Fdocs\u002Fguides\u002Fagents), [Google ADK](https:\u002F\u002Fgoogle.github.io\u002Fadk-docs\u002F), and [LangChain Deep Agents](https:\u002F\u002Fdocs.langchain.com\u002Foss\u002Fpython\u002Fdeepagents\u002Foverview)\n- **Open protocols and interoperability**, including [Model Context Protocol](https:\u002F\u002Fmodelcontextprotocol.io\u002Fspecification\u002F2025-06-18), [A2A](https:\u002F\u002Fa2a-protocol.org\u002Flatest\u002F), [AG-UI](https:\u002F\u002Fdocs.ag-ui.com\u002F), and [AgentSchema](https:\u002F\u002Fmicrosoft.github.io\u002FAgentSchema\u002F)\n- **Coding agents and project memory**, including [OpenAI Codex](https:\u002F\u002Fopenai.com\u002Findex\u002Fintroducing-codex\u002F), [Claude Code memory](https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fdocs\u002Fclaude-code\u002Fmemory), and [Letta memory blocks](https:\u002F\u002Fdocs.letta.com\u002Fguides\u002Fcore-concepts\u002Fmemory\u002Fmemory-blocks)\n- **Evaluation and observability**, including [LangSmith observability](https:\u002F\u002Fdocs.langchain.com\u002Flangsmith\u002Fobservability-quickstart) and [OpenTelemetry semantic conventions for GenAI](https:\u002F\u002Fopentelemetry.io\u002Fdocs\u002Fspecs\u002Fsemconv\u002Fgen-ai\u002F)\n\n---\n\n## 📚 Table of Contents\n\n- [Awesome Context Engineering](#awesome-context-engineering)\n  - [💬 Join Our Community](#-join-our-community)\n  - [📧 Contact](#-contact)\n  - [📰 News](#-news)\n  - [🎯 Introduction](#-introduction)\n  - [🧭 2026 Agent Era Update](#-2026-agent-era-update)\n    - [From Context Engineering to Agent Engineering](#from-context-engineering-to-agent-engineering)\n    - [What This Repository Now Covers](#what-this-repository-now-covers)\n    - [Reading Guide for 2026 Topics](#reading-guide-for-2026-topics)\n  - [📚 Table of Contents](#-table-of-contents)\n  - [🔗 Related Survey](#-related-survey)\n  - [🏗️ Definition of Context Engineering](#️-definition-of-context-engineering)\n    - [LLM Generation](#llm-generation)\n    - [Definition of Context](#definition-of-context)\n    - [Definition of Context Engineering](#definition-of-context-engineering)\n    - [Dynamic Context Orchestration](#dynamic-context-orchestration)\n    - [Mathematical Principles](#mathematical-principles)\n    - [Theoretical Framework: Bayesian Context Inference](#theoretical-framework-bayesian-context-inference)\n    - [Comparison](#comparison)\n  - [🌐 Related Blogs](#-related-blogs)\n    - [Social Media \\& Talks](#social-media--talks)\n  - [🤔 Why Context Engineering?](#-why-context-engineering)\n    - [The Paradigm Shift: From Tactical to Strategic](#the-paradigm-shift-from-tactical-to-strategic)\n    - [1. Fundamental Challenges with Current Approaches](#1-fundamental-challenges-with-current-approaches)\n      - [Human Intent Communication Challenges](#human-intent-communication-challenges)\n      - [Complex Knowledge Requirements](#complex-knowledge-requirements)\n      - [Reliability and Trustworthiness Issues](#reliability-and-trustworthiness-issues)\n    - [2. Limitations of Static Prompting](#2-limitations-of-static-prompting)\n      - [From Strings to Systems](#from-strings-to-systems)\n      - [The \"Movie Production\" Analogy](#the-movie-production-analogy)\n    - [3. Enterprise and Production Requirements](#3-enterprise-and-production-requirements)\n      - [Context Failures Are the New Bottleneck](#context-failures-are-the-new-bottleneck)\n      - [Scalability Beyond Simple Tasks](#scalability-beyond-simple-tasks)\n      - [Reliability and Consistency](#reliability-and-consistency)\n      - [Economic and Operational Efficiency](#economic-and-operational-efficiency)\n    - [4. Cognitive and Information Science Foundations](#4-cognitive-and-information-science-foundations)\n      - [Artificial Embodiment](#artificial-embodiment)\n      - [Information Retrieval at Scale](#information-retrieval-at-scale)\n    - [5. The Future of AI System Architecture](#5-the-future-of-ai-system-architecture)\n  - [🔧 Components, Techniques and Architectures](#-components-techniques-and-architectures)\n    - [Context Scaling](#context-scaling)\n    - [Context Management in Production](#context-management-in-production)\n    - [Structured Data Integration](#structured-data-integration)\n    - [Self-Generated Context](#self-generated-context)\n  - [🛠️ Implementation and Challenges](#️-implementation-and-challenges)\n    - [0. Agent Harnesses and Runtime Systems](#0-agent-harnesses-and-runtime-systems)\n    - [1. Retrieval-Augmented Generation (RAG)](#1-retrieval-augmented-generation-rag)\n    - [2. Memory Systems](#2-memory-systems)\n      - [Runtime Memory Design Patterns](#runtime-memory-design-patterns)\n      - [Project Memory and Instruction Artifacts](#project-memory-and-instruction-artifacts)\n    - [3. Agent Communication](#3-agent-communication)\n      - [Open Agent Protocols and Interoperability](#open-agent-protocols-and-interoperability)\n    - [4. Tool Use and Function Calling](#4-tool-use-and-function-calling)\n      - [Hosted Agent Tools and Computer Use](#hosted-agent-tools-and-computer-use)\n  - [📊 Evaluation Paradigms for Context-Driven Systems](#-evaluation-paradigms-for-context-driven-systems)\n    - [Context Quality Assessment](#context-quality-assessment)\n    - [Benchmarking Context Engineering](#benchmarking-context-engineering)\n    - [Agent Observability and Telemetry](#agent-observability-and-telemetry)\n  - [🚀 Applications and Systems](#-applications-and-systems)\n    - [Complex Research Systems](#complex-research-systems)\n    - [Production Systems](#production-systems)\n      - [Coding Agents and Project Memory](#coding-agents-and-project-memory)\n      - [Platform Stacks and Hosted Agent Runtimes](#platform-stacks-and-hosted-agent-runtimes)\n  - [🔮 Limitations and Future Directions](#-limitations-and-future-directions)\n    - [Current Limitations](#current-limitations)\n    - [Future Research Directions](#future-research-directions)\n  - [🤝 Contributing](#-contributing)\n    - [Paper Formatting Guidelines](#paper-formatting-guidelines)\n    - [Badge Colors](#badge-colors)\n  - [📄 License](#-license)\n  - [📑 Citation](#-citation)\n  - [⚠️ Disclaimer](#️-disclaimer)\n  - [📧 Contact](#-contact-1)\n  - [🙏 Acknowledgments](#-acknowledgments)\n  - [Star History](#star-history)\n  - [📖 Our Paper](#-our-paper)\n\n---\n\n## 🔗 Related Survey\n\n\u003Cb>General AI Survey Papers\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>A Survey of Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Zhao et al.,\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2303.18223\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.03-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FRUCAIBox\u002FLLMSurvey\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FRUCAIBox\u002FLLMSurvey.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>The Prompt Report: A Systematic Survey of Prompt Engineering Techniques\u003C\u002Fb>\u003C\u002Fi>, Schulhoff et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2406.06608\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Ftrigaten\u002FThe_Prompt_Report\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Ftrigaten\u002FThe_Prompt_Report.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>A Systematic Survey of Prompt Engineering in Large Language Models: Techniques and Applications\u003C\u002Fb>\u003C\u002Fi>, Sahoo et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.07927\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.03-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>A Systematic Survey of Prompt Engineering on Vision-Language Foundation Models\u003C\u002Fb>\u003C\u002Fi>, Gao et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2307.12980\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.07-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FJindongGu\u002FAwesome-Prompting-on-Vision-Language-Model\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FJindongGu\u002FAwesome-Prompting-on-Vision-Language-Model.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Context and Reasoning\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>A Survey on In-context Learning\u003C\u002Fb>\u003C\u002Fi>, Dong et al., \u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.18653\u002Fv1\u002F2024.emnlp-main.64\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FEMNLP-2024.11-blue\" alt=\"EMNLP Badge\">\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fdqxiu\u002FICL_PaperList\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fdqxiu\u002FICL_PaperList.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>The Mystery of In-Context Learning: A Comprehensive Survey on Interpretation and Analysis\u003C\u002Fb>\u003C\u002Fi>, Zhou et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2311.00237\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.10-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fzyxnlp\u002FICL-Interpretation-Analysis-Resources\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fzyxnlp\u002FICL-Interpretation-Analysis-Resources.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>A Comprehensive Survey of Retrieval-Augmented Generation (RAG): Evolution, Current Landscape and Future Directions\u003C\u002Fb>\u003C\u002Fi>, Gupta et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.12837\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.10-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Retrieval-Augmented Generation for Large Language Models: A Survey\u003C\u002Fb>\u003C\u002Fi>, Gao et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.10997\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.03-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FTongji-KGLLM\u002FRAG-Survey\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FTongji-KGLLM\u002FRAG-Survey.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>A Survey on Knowledge-Oriented Retrieval-Augmented Generation\u003C\u002Fb>\u003C\u002Fi>, Cheng et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2503.10677\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FUSTCAGI\u002FAwesome-Papers-Retrieval-Augmented-Generation\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FUSTCAGI\u002FAwesome-Papers-Retrieval-Augmented-Generation.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Memory Systems and Context Persistence\u003C\u002Fb>\n\n\u003Cb>Survey\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>A Survey on the Memory Mechanism of Large Language Model based Agents\u003C\u002Fb>\u003C\u002Fi>, Zhang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.13501\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fnuster1128\u002FLLM_Agent_Memory_Survey\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fnuster1128\u002FLLM_Agent_Memory_Survey.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Survey on Memory-Augmented Neural Networks: Cognitive Insights to AI Applications\u003C\u002Fb>\u003C\u002Fi>, Khosla et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.06141\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.12-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>From Human Memory to AI Memory: A Survey on Memory Mechanisms in the Era of LLMs\u003C\u002Fb>\u003C\u002Fi>, Wu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2504.15965\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Survey on Evaluation of LLM-based Agents\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2503.16416\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.03-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>A Survey of Personalized Large Language Models: Progress and Future Directions\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.11528\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Agentic Retrieval-Augmented Generation: A Survey\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.09136\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.01-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Retrieval-Augmented Generation with Graphs (GraphRAG)\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.00309\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.12-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FGraph-RAG\u002FGraphRAG\u002F\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FGraph-RAG\u002FGraphRAG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>The Landscape of Agentic Reinforcement Learning for LLMs: A Survey\u003C\u002Fb>\u003C\u002Fi>, Zhang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2509.02547\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.09-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fxhyumiracle\u002FAwesome-AgenticLLM-RL-Papers\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fxhyumiracle\u002FAwesome-AgenticLLM-RL-Papers.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Benchmarks\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Evaluating Very Long-Term Conversational Memory of LLM Agents (LOCOMO)\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.17753\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FACL-2024.02-blue\" alt=\"ACL Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fsnap-research.github.io\u002Flocomo\u002F\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fsnap-research\u002Flocomo.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions\u003C\u002Fb>\u003C\u002Fi>, Hu et al.,\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.05257\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.07-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FHUST-AI-HYZ\u002FMemoryAgentBench\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FHUST-AI-HYZ\u002FMemoryAgentBench.svg?style=social\" alt=\"GitHub stars\">\u003C\u002Fa>\n      \u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fdatasets\u002Fai-hyz\u002FMemoryAgentBench\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fhuggingface.co\u002Fdatasets\u002Fhuggingface\u002Fbadges\u002Fresolve\u002Fmain\u002Fdataset-on-hf-sm.svg\" alt=\"HF Dataset\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Episodic Memories Generation and Evaluation Benchmark for Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.13121\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.01-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>On the Structural Memory of LLM Agents\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.15266\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.12-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering\u003C\u002Fb>\u003C\u002Fi>, Yang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F1809.09600\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FEMNLP-2018.09-blue\" alt=\"EMNLP Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fhotpotqa.github.io\u002F\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fhotpotqa\u002Fhotpot.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cb>Neural Memory Architectures\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Neural Turing Machines\u003C\u002Fb>\u003C\u002Fi>, Graves et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F1410.5401\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2014.10-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Differentiable Neural Computers\u003C\u002Fb>\u003C\u002Fi>, Graves et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F1610.06258\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2016.10-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fgoogle-deepmind\u002Fdnc\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fgoogle-deepmind\u002Fdnc.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>A Brain-inspired Memory Transformation based Differentiable Neural Computer\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2301.02809\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.01-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Differentiable Neural Computers with Memory Demon\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2211.02987\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.11-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cb>Memory-Augmented Transformers\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Memorizing Transformers\u003C\u002Fb>\u003C\u002Fi>, Wu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2203.08913\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.03-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Recurrent Memory Transformer\u003C\u002Fb>\u003C\u002Fi>, Bulatov et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2207.06881\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNeurIPS-2022.07-blue\" alt=\"NeurIPS Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fbooydar\u002Frecurrent-memory-transformer\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fbooydar\u002Frecurrent-memory-transformer.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Leave No Context Behind: Efficient Infinite Context Transformers with Infini-attention\u003C\u002Fb>\u003C\u002Fi>, Munkhdalai et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.07143\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Memformer: A Memory-Augmented Transformer for Sequence Modeling\u003C\u002Fb>\u003C\u002Fi>, Wu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2010.06891\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2020.10-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Token Turing Machines\u003C\u002Fb>\u003C\u002Fi>, Ryoo et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2211.09119\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.11-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>TransformerFAM: Feedback Attention is Working Memory\u003C\u002Fb>\u003C\u002Fi>, Irie et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.09173\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Production Memory Systems\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>MemGPT: Towards LLMs as Operating Systems\u003C\u002Fb>\u003C\u002Fi>, Packer et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.08560\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fresearch.memgpt.ai\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fletta-ai\u002Fletta.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>MemoryBank: Enhancing Large Language Models with Long-Term Memory\u003C\u002Fb>\u003C\u002Fi>, Zhong et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2305.10250\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fzhongwanjun\u002Fmemorybank-siliconfriend\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fzhongwanjun\u002Fmemorybank-siliconfriend.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>MEM0: Building Production-Ready AI Agents with Scalable Long-Term Memory\u003C\u002Fb>\u003C\u002Fi>, Taranjeet et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2504.19413\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fmem0.ai\u002Fresearch\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmem0ai\u002Fmem0.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>MEM1: Learning to Synergize Memory and Reasoning for Efficient Long-Horizon Agents\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2506.15841\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.06-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmannaandpoem\u002Fopenmanus\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmannaandpoem\u002Fopenmanus.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>A-MEM: Agentic Memory for LLM Agents\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.12110\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fagiresearch\u002FA-mem\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fagiresearch\u002FA-mem.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>MemAgent: Reshaping Long-Context LLM with Multi-Conv RL-based Memory Agent\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.02259\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.07-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Memory OS of AI Agent\u003C\u002Fb>\u003C\u002Fi>, Kang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2506.06326\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FBAI-LAB\u002FMemoryOS\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FBAI-LAB\u002FMemoryOS.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cb>Graph-based Memory Systems\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>arigraph: learning knowledge graph world models with episodic memory for llm agents\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.04363\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.07-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Zep: A Temporal Knowledge Graph Architecture for Agent Memory\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.13956\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.01-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fgetzep\u002Fgraphiti\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fgetzep\u002Fgraphiti.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>KG-Agent: An Efficient Autonomous Agent Framework for Complex Reasoning over Knowledge Graph\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.11163\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>GraphReader: Building Graph-based Agent to Enhance Long-Context Abilities of Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2406.14550\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.06-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>From Local to Global: A GraphRAG Approach to Query-Focused Summarization\u003C\u002Fb>\u003C\u002Fi>, Edge et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.16130\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002Fgraphrag\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmicrosoft\u002Fgraphrag.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Knowledge Graph-Guided Retrieval Augmented Generation\u003C\u002Fb>\u003C\u002Fi>, Zhu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.06864\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cb>Episodic and Working Memory\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Larimar: Large Language Models with Episodic Memory Control\u003C\u002Fb>\u003C\u002Fi>, Goyal et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2403.11901\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICML-2024.03-blue\" alt=\"ICML Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>EM-LLM: Human-like Episodic Memory for Infinite Context LLMs\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.09450\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2024.07-blue\" alt=\"ICLR Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fem-llm\u002FEM-LLM-model\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fem-llm\u002FEM-LLM-model.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Large Language Models with Controllable Working Memory\u003C\u002Fb>\u003C\u002Fi>, Goyal et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2211.05110\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.11-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Empowering Working Memory for Large Language Model Agents\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.17259\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.12-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cb>Conversational Memory\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>MemoChat: Tuning LLMs to Use Memos for Consistent Long-Range Open-Domain Conversation\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2308.08239\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.08-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Think-in-Memory: Recalling and Post-thinking Enable LLMs with Long-Term Memory\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2311.08719\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.11-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Generative Agents: Interactive Simulacra of Human Behavior\u003C\u002Fb>\u003C\u002Fi>, Park et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2304.03442\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Self-Controlled Memory Framework for Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2304.13343\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cb>Foundational Survey Papers from Major Venues\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>AUTOPROMPT: Eliciting Knowledge from Language Models with Automatically Generated Prompts\u003C\u002Fb>\u003C\u002Fi>, Shin et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FEMNLP-2020-blue\" alt=\"EMNLP Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fucinlp\u002Fautoprompt\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fucinlp\u002Fautoprompt.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>The Power of Scale for Parameter-Efficient Prompt Tuning\u003C\u002Fb>\u003C\u002Fi>, Lester et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FEMNLP-2021-blue\" alt=\"EMNLP Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fgoogle-research\u002Fprompt-tuning\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fgoogle-research\u002Fprompt-tuning.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Prefix-Tuning: Optimizing Continuous Prompts for Generation\u003C\u002Fb>\u003C\u002Fi>, Li et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FACL-2021-blue\" alt=\"ACL Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FXiangLi1999\u002FPrefixTuning\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FXiangLi1999\u002FPrefixTuning.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>An Explanation of In-context Learning as Implicit Bayesian Inference\u003C\u002Fb>\u003C\u002Fi>, Xie et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2022-blue\" alt=\"ICLR Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fp-lambda\u002Fincontext-learning\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fp-lambda\u002Fincontext-learning.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Rethinking the Role of Demonstrations: What Makes In-context Learning Work?\u003C\u002Fb>\u003C\u002Fi>, Min et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FEMNLP-2022-blue\" alt=\"EMNLP Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FAlrope123\u002Frethinking-demonstrations\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FAlrope123\u002Frethinking-demonstrations.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\u003Cb>Additional RAG and Retrieval Surveys\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Retrieval-Augmented Generation for AI-Generated Content: A Survey\u003C\u002Fb>\u003C\u002Fi>, Various, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.19473\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FPKU-DAIR\u002FRAG-Survey\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FPKU-DAIR\u002FRAG-Survey.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Retrieval Augmented Generation (RAG) and Beyond: A Comprehensive Survey on How to Make your LLMs use External Data More Wisely\u003C\u002Fb>\u003C\u002Fi>, Various, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2409.14924\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.09-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Large language models (LLMs): survey, technical frameworks, and future challenges\u003C\u002Fb>\u003C\u002Fi>, Various, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAIR-2024-blue\" alt=\"AIR Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\n---\n\n## 🏗️ Definition of Context Engineering\n\n> **Context is not just the single prompt users send to an LLM. Context is the complete information payload provided to a LLM at inference time, encompassing all structured informational components that the model needs to plausibly accomplish a given task.**\n\n### LLM Generation\n\nTo formally define Context Engineering, we must first mathematically characterize the LLM generation process. Let us model an LLM as a probabilistic function:\n\n$$P(\\text{output} | \\text{context}) = \\prod_{t=1}^T P(\\text{token}_t | \\text{previous tokens}, \\text{context})$$\n\nWhere:\n- $\\text{context}$ represents the complete input information provided to the LLM\n- $\\text{output}$ represents the generated response sequence\n- $P(\\text{token}_t | \\text{previous tokens}, \\text{context})$ is the probability of generating each token given the context\n\n### Definition of Context\n\nIn traditional prompt engineering, the context is treated as a simple string:\n$$\\text{context} = \\text{prompt}$$\n\nHowever, in Context Engineering, we decompose the context into multiple structured components:\n\n$$\\text{context} = \\text{Assemble}(\\text{instructions}, \\text{knowledge}, \\text{tools}, \\text{memory}, \\text{state}, \\text{query})$$\n\nWhere $\\text{Assemble}$ is a context assembly function that orchestrates:\n- $\\text{instructions}$: System prompts and rules\n- $\\text{knowledge}$: Retrieved relevant information\n- $\\text{tools}$: Available function definitions\n- $\\text{memory}$: Conversation history and learned facts\n- $\\text{state}$: Current world\u002Fuser state\n- $\\text{query}$: User's immediate request\n\n### Definition of Context Engineering\n\n**Context Engineering** is formally defined as the optimization problem:\n\n$$\\text{Assemble}^* = \\arg\\max_{\\text{Assemble}} \\mathbb{E} [\\text{Reward}(\\text{LLM}(\\text{context}), \\text{target})]$$\n\nSubject to constraints:\n- $|\\text{context}| \\leq \\text{MaxTokens} \\text{(context window limitation)}$\n- $\\text{knowledge} = \\text{Retrieve}(\\text{query}, \\text{database})$\n- $\\text{memory} = \\text{Select}(\\text{history}, \\text{query})$\n- $\\text{state} = \\text{Extract}(\\text{world})$\n\nWhere:\n- $\\text{Reward}$ measures the quality of generated responses\n- $\\text{Retrieve}$, $\\text{Select}$, $\\text{Extract}$ are functions for information gathering\n\n### Dynamic Context Orchestration\n\nThe context assembly can be decomposed as:\n\n$$\\text{context} = \\text{Concat}(\\text{Format}(\\text{instructions}), \\text{Format}(\\text{knowledge}), \\text{Format}(\\text{tools}), \\text{Format}(\\text{memory}), \\text{Format}(\\text{query}))$$\n\nWhere $\\text{Format}$ represents component-specific structuring, and $\\text{Concat}$ assembles them respecting token limits and optimal positioning.\n\n**Context Engineering** is therefore the discipline of designing and optimizing these assembly and formatting functions to maximize task performance.\n\n### Mathematical Principles\n\nFrom this formalization, we derive four fundamental principles:\n\n1. **System-Level Optimization**: Context generation is a multi-objective optimization problem over assembly functions, not simple string manipulation.\n\n2. **Dynamic Adaptation**: The context assembly function adapts to each $\\text{query}$ and $\\text{state}$ at inference time: $\\text{Assemble}(\\cdot | \\text{query}, \\text{state})$.\n\n3. **Information-Theoretic Optimality**: The retrieval function maximizes relevant information: $\\text{Retrieve} = \\arg\\max \\text{Relevance}(\\text{knowledge}, \\text{query})$.\n\n4. **Structural Sensitivity**: The formatting functions encode structure that aligns with LLM processing capabilities.\n\n### Theoretical Framework: Bayesian Context Inference\n\nContext Engineering can be formalized within a Bayesian framework where the optimal context is inferred:\n\n$$P(\\text{context} | \\text{query}, \\text{history}, \\text{world}) \\propto P(\\text{query} | \\text{context}) \\cdot P(\\text{context} | \\text{history}, \\text{world})$$\n\nWhere:\n- $P(\\text{query} | \\text{context})$ models query-context compatibility\n- $P(\\text{context} | \\text{history}, \\text{world})$ represents prior context probability\n\nThe optimal context assembly becomes:\n\n$$\\text{context}^* = \\arg\\max_{\\text{context}} P(\\text{answer} | \\text{query}, \\text{context}) \\cdot P(\\text{context} | \\text{query}, \\text{history}, \\text{world})$$\n\nThis Bayesian formulation enables:\n- **Uncertainty Quantification**: Modeling confidence in context relevance\n- **Adaptive Retrieval**: Updating context beliefs based on feedback\n- **Multi-step Reasoning**: Maintaining context distributions across interactions\n\n### Comparison\n\n| Dimension | Prompt Engineering | Context Engineering |\n|-----------|-------------------|-------------------|\n| **Mathematical Model** | $\\text{context} = \\text{prompt}$ (static) | $\\text{context} = \\text{Assemble}(...)$ (dynamic) |\n| **Optimization Target** | $\\arg\\max_{\\text{prompt}} P(\\text{answer} \\mid \\text{query}, \\text{prompt})$ | $\\arg\\max_{\\text{Assemble}} \\mathbb{E}[\\text{Reward}(...)]$ |\n| **Complexity** | $O(1)$ context assembly | $O(n)$ multi-component optimization |\n| **Information Theory** | Fixed information content | Adaptive information maximization |\n| **State Management** | Stateless function | Stateful with $\\text{memory}(\\text{history}, \\text{query})$ |\n| **Scalability** | Linear in prompt length | Sublinear through compression\u002Ffiltering |\n| **Error Analysis** | Manual prompt inspection | Systematic evaluation of assembly components |\n\n\n\n---\n\n## 🌐 Related Blogs\n\n- [The rise of \"context engineering\"](https:\u002F\u002Fblog.langchain.com\u002Fthe-rise-of-context-engineering\u002F)\n- [The New Skill in AI is Not Prompting, It's Context Engineering](https:\u002F\u002Fwww.philschmid.de\u002Fcontext-engineering)\n- [davidkimai\u002FContext-Engineering: \"Context engineering is the delicate art and science of filling the context window with just the right information for the next step.\" ](https:\u002F\u002Fgithub.com\u002Fdavidkimai\u002FContext-Engineering)\n- [Context Engineering is Runtime of AI Agents | by Bijit Ghosh | Jun, 2025 | Medium](https:\u002F\u002Fmedium.com\u002F@bijit211987\u002Fcontext-engineering-is-runtime-of-ai-agents-411c9b2ef1cb)\n- [Context Engineering](https:\u002F\u002Fblog.langchain.com\u002Fcontext-engineering-for-agents\u002F)\n- [Context Engineering for Agents](https:\u002F\u002Frlancemartin.github.io\u002F2025\u002F06\u002F23\u002Fcontext_engineering\u002F)\n- [Cognition | Don't Build Multi-Agents](https:\u002F\u002Fcognition.ai\u002Fblog\u002Fdont-build-multi-agents)\n- [从Prompt Engineering到Context Engineering - 53AI-AI知识库|大模型知识库|大模型训练|智能体开发](https:\u002F\u002Fwww.53ai.com\u002Fnews\u002Ftishicikuangjia\u002F2025062727685.html)\n\n### Social Media & Talks\n\n- [Mastering Claude Code in 30 minutes](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=6eBSHbLKuN0)\n- [Context Engineering for Agents](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=4GiqzUHD5AA)\n- [Andrej Karpathy on X: \"+1 for \"context engineering\" over \"prompt engineering\"](https:\u002F\u002Fx.com\u002Fkarpathy\u002Fstatus\u002F1937902205765607626?ref=blog.langchain.com)\n- [复旦大学\u002F上海创智学院邱锡鹏：Context Scaling，通往AGI的下一幕](https:\u002F\u002Fmp.weixin.qq.com\u002Fs\u002FKnej0qbyr5j5KX_BO7FGew)\n\n---\n\n## 🤔 Why Context Engineering?\n\n### The Paradigm Shift: From Tactical to Strategic\n\nThe evolution from prompt engineering to context engineering represents a fundamental maturation in AI system design. As influential figures like Andrej Karpathy, Tobi Lutke, and Simon Willison have argued, the term \"prompt engineering\" has been diluted to mean simply \"typing things into a chatbot,\" failing to capture the complexity required for industrial-strength LLM applications.\n\n### 1. Fundamental Challenges with Current Approaches\n\n#### Human Intent Communication Challenges\n- **Unclear Human Intent Expression**: Human intentions are often unclear, incomplete, or ambiguous when expressed in natural language\n- **AI's Incomplete Understanding of Human Intent**: AI systems struggle to fully comprehend complex human intentions, especially those involving implicit context or cultural nuances\n- **Overly Literal AI Interpretation**: AI systems often interpret human instructions too literally, missing the underlying intent or contextual meaning\n\n#### Complex Knowledge Requirements\nSingle models alone cannot solve complex problems that require:\n- **(1) Large-scale External Knowledge**: Vast amounts of external knowledge that exceed model capacity\n- **(2) Accurate External Knowledge**: Precise, up-to-date information that models may not possess\n- **(3) Novel External Knowledge**: Emerging knowledge that appears after model training\n\n**Static Knowledge Limitations:**\n- **Static Knowledge Problem**: Pre-trained models contain static knowledge that becomes outdated\n- **Knowledge Cutoff**: Models cannot access information beyond their training data\n- **Domain-Specific Gaps**: Models lack specialized knowledge for specific industries or applications\n\n#### Reliability and Trustworthiness Issues\n- **AI Hallucination**: LLMs generate plausible but factually incorrect information when lacking proper context\n- **Lack of Provenance**: Absence of clear source attribution for generated information\n- **Confidence Calibration**: Models often appear confident even when generating false information\n- **Transparency Gaps**: Inability to trace how conclusions were reached\n- **Accountability Issues**: Difficulty in verifying the reliability of AI-generated content\n\n### 2. Limitations of Static Prompting\n\n#### From Strings to Systems\nTraditional prompting treats context as a static string, but enterprise applications require:\n- **Dynamic Information Assembly**: Context created on-the-fly, tailored to specific users and queries\n- **Multi-Source Integration**: Combining databases, APIs, documents, and real-time data\n- **State Management**: Maintaining conversation history, user preferences, and workflow status\n- **Tool Orchestration**: Coordinating external function calls and API interactions\n\n#### The \"Movie Production\" Analogy\nIf prompt engineering is writing a single line of dialogue for an actor, context engineering is the entire process of building the set, designing lighting, providing detailed backstory, and directing the scene. The dialogue only achieves its intended impact because of the rich, carefully constructed environment surrounding it.\n\n### 3. Enterprise and Production Requirements\n\n#### Context Failures Are the New Bottleneck\nMost failures in modern agentic systems are no longer attributable to core model reasoning capabilities but are instead **\"context failures\"**. The true engineering challenge lies not in what question to ask, but in ensuring the model has all necessary background, data, tools, and memory to answer meaningfully and reliably.\n\n#### Scalability Beyond Simple Tasks\nWhile prompt engineering suffices for simple, self-contained tasks, it breaks down when scaled to:\n- **Complex, multi-step applications**\n- **Data-rich enterprise environments** \n- **Stateful, long-running workflows**\n- **Multi-user, multi-tenant systems**\n\n#### Reliability and Consistency\nEnterprise applications demand:\n- **Deterministic Behavior**: Predictable outputs across different contexts and users\n- **Error Handling**: Graceful degradation when information is incomplete or contradictory\n- **Audit Trails**: Transparency in how context influences model decisions\n- **Compliance**: Meeting regulatory requirements for data handling and decision making\n\n#### Economic and Operational Efficiency\nContext Engineering enables:\n- **Cost Optimization**: Strategic choice between RAG and long-context approaches\n- **Latency Management**: Efficient information retrieval and context assembly\n- **Resource Utilization**: Optimal use of finite context windows and computational resources\n- **Maintenance Scalability**: Systematic approaches to updating and managing knowledge bases\n\nContext Engineering provides the architectural foundation for managing state, integrating diverse data sources, and maintaining coherence across these demanding scenarios.\n\n### 4. Cognitive and Information Science Foundations\n\n#### Artificial Embodiment\nLLMs are essentially \"brains in a vat\" - powerful reasoning engines lacking connection to specific environments. Context Engineering provides:\n- **Synthetic Sensory Systems**: Retrieval mechanisms as artificial perception\n- **Proxy Embodiment**: Tool use as artificial action capabilities  \n- **Artificial Memory**: Structured information storage and retrieval\n\n#### Information Retrieval at Scale\nContext Engineering addresses the fundamental challenge of information retrieval where the \"user\" is not human but an AI agent. This requires:\n- **Semantic Understanding**: Bridging the gap between intent and expression\n- **Relevance Optimization**: Ranking and filtering vast knowledge bases\n- **Query Transformation**: Converting ambiguous requests into precise retrieval operations\n\n### 5. The Future of AI System Architecture\n\nContext Engineering elevates AI development from a collection of \"prompting tricks\" to a rigorous discipline of systems architecture. It applies decades of knowledge in operating system design, memory management, and distributed systems to the unique challenges of LLM-based applications.\n\nThis discipline is foundational for unlocking the full potential of LLMs in production systems, enabling the transition from one-off text generation to autonomous agents and sophisticated AI copilots that can reliably operate in complex, dynamic environments.\n\n---\n\n## 🔧 Components, Techniques and Architectures\n\n### Context Scaling\n\n\u003Cb>Position Interpolation and Extension Techniques\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Extending Context Window of Large Language Models via Position Interpolation\u003C\u002Fb>\u003C\u002Fi>, Chen et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2306.15595\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.06-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FMath1019\u002FExtend_Context_Window_Position_Interpolation\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FMath1019\u002FExtend_Context_Window_Position_Interpolation.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>YaRN: Efficient Context Window Extension of Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Peng et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2309.00071\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2024.01-blue\" alt=\"ICLR Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fjquesnelle\u002Fyarn\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fjquesnelle\u002Fyarn.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LongRoPE: Extending LLM Context Window Beyond 2 Million Tokens\u003C\u002Fb>\u003C\u002Fi>, Ding et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.13753\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICML-2024.02-blue\" alt=\"ICML Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002FLongRoPE\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmicrosoft\u002FLongRoPE.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LongRoPE2: Near-Lossless LLM Context Window Scaling\u003C\u002Fb>\u003C\u002Fi>, Shang et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICML-2025.05-blue\" alt=\"ICML Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002FLongRoPE\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmicrosoft\u002FLongRoPE.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Memory-Efficient Attention Mechanisms\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Fast Multipole Attention: A Divide-and-Conquer Attention Mechanism for Long Sequences\u003C\u002Fb>\u003C\u002Fi>, Kang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.11960\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2024.02-blue\" alt=\"ICLR Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fyanmingk\u002FFMA\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fyanmingk\u002FFMA.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Leave No Context Behind: Efficient Infinite Context Transformers with Infini-attention\u003C\u002Fb>\u003C\u002Fi>, Munkhdalai et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.07143\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fjlamprou\u002FInfini-Attention\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fjlamprou\u002FInfini-Attention.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Heads\u003C\u002Fb>\u003C\u002Fi>, Xiao et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2025.01-blue\" alt=\"ICLR Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmit-han-lab\u002Fduo-attention\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmit-han-lab\u002Fduo-attention.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Star Attention: Efficient LLM Inference over Long Sequences\u003C\u002Fb>\u003C\u002Fi>, Acharya et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2411.17116\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.11-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FStar-Attention\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FNVIDIA\u002FStar-Attention.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Ultra-Long Sequence Processing (100K+ Tokens)\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>TokenSwift: Lossless Acceleration of Ultra Long Sequence Generation\u003C\u002Fb>\u003C\u002Fi>, Wu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.18890\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICML-2025.02-blue\" alt=\"ICML Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fbigai-nlco\u002FTokenSwift\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fbigai-nlco\u002FTokenSwift.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LongHeads: Multi-Head Attention is Secretly a Long Context Processor\u003C\u002Fb>\u003C\u002Fi>, Lu et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FEMNLP-2024.11-blue\" alt=\"EMNLP Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FLuLuLuyi\u002FLongHeads\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FLuLuLuyi\u002FLongHeads.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>∞Bench: Extending Long Context Evaluation Beyond 100K Tokens\u003C\u002Fb>\u003C\u002Fi>, Bai et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.00359\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FACL-2024.06-blue\" alt=\"ACL Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FOpenBMB\u002FInfiniteBench\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FOpenBMB\u002FInfiniteBench.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Comprehensive Extension Surveys and Methods\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Beyond the Limits: A Survey of Techniques to Extend the Context Length in Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Various, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.02244\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>A Controlled Study on Long Context Extension and Generalization in LLMs\u003C\u002Fb>\u003C\u002Fi>, Various, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2409.12181\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.09-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FLeooyii\u002FLCEG\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FLeooyii\u002FLCEG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Selective Attention: Enhancing Transformer through Principled Context Control\u003C\u002Fb>\u003C\u002Fi>, Various, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNeurIPS-2024-blue\" alt=\"NeurIPS Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fumich-sota\u002Fselective_attention\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fumich-sota\u002Fselective_attention.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cb>Vision-Language Models with Sophisticated Context Understanding\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Towards LLM-Centric Multimodal Fusion: A Survey on Integration Strategies and Techniques\u003C\u002Fb>\u003C\u002Fi>, An et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2506.04788\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.01-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Browse and Concentrate: Comprehending Multimodal Content via Prior-LLM Context Fusion\u003C\u002Fb>\u003C\u002Fi>, Wang et al., \u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.18653\u002Fv1\u002F2024.acl-long.605\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FACL-2024.08-blue\" alt=\"ACL Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FTHUNLP-MT\u002FBrote\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FTHUNLP-MT\u002FBrote.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>V2PE: Improving Multimodal Long-Context Capability of Vision-Language Models with Variable Visual Position Encoding\u003C\u002Fb>\u003C\u002Fi>, Dai et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.09616\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.12-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FOpenGVLab\u002FV2PE\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FOpenGVLab\u002FV2PE.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Flamingo: a Visual Language Model for Few-Shot Learning\u003C\u002Fb>\u003C\u002Fi>, Alayrac et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2204.14198\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNeurIPS-2022.04-blue\" alt=\"NeurIPS Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Flucidrains\u002Fflamingo-pytorch\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Flucidrains\u002Fflamingo-pytorch.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n\u003Cb>Audio-Visual Context Integration and Processing\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Aligned Better, Listen Better for Audio-Visual Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Guo et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2025.01-blue\" alt=\"ICLR Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>AVicuna: Audio-Visual LLM with Interleaver and Context-Boundary Alignment for Temporal Referential Dialogue\u003C\u002Fb>\u003C\u002Fi>, Chen et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2403.16276\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.03-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>SonicVisionLM: Playing Sound with Vision Language Models\u003C\u002Fb>\u003C\u002Fi>, Xie et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.04394\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FCVPR-2024.01-blue\" alt=\"CVPR Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FYusiissy\u002FSonicVisionLM\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FYusiissy\u002FSonicVisionLM.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>SAVEn-Vid: Synergistic Audio-Visual Integration for Enhanced Understanding in Long Video Context\u003C\u002Fb>\u003C\u002Fi>, Li et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2411.16213\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.11-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FLJungang\u002FSAVEn-Vid\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FLJungang\u002FSAVEn-Vid.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n\n\u003Cb>Multi-Modal Prompt Engineering and Context Design\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>CaMML: Context-Aware Multimodal Learner for Large Models\u003C\u002Fb>\u003C\u002Fi>, Chen et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.11406\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FACL-2024.08-blue\" alt=\"ACL Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Visual In-Context Learning for Large Vision-Language Models\u003C\u002Fb>\u003C\u002Fi>, Zhou et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FACL-2024.08-blue\" alt=\"ACL Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>CAMA: Enhancing Multimodal In-Context Learning with Context-Aware Modulated Attention\u003C\u002Fb>\u003C\u002Fi>, Li et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2505.17097\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n\n\u003Cb>CVPR 2024 Vision-Language Advances\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>CogAgent: A Visual Language Model for GUI Agents\u003C\u002Fb>\u003C\u002Fi>, Various, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FCVPR-2024-blue\" alt=\"CVPR Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FTHUDM\u002FCogAgent\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FTHUDM\u002FCogAgent.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LISA: Reasoning Segmentation via Large Language Model\u003C\u002Fb>\u003C\u002Fi>, Various, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FCVPR-2024-blue\" alt=\"CVPR Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fdvlab-research\u002FLISA\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fdvlab-research\u002FLISA.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Reproducible scaling laws for contrastive language-image learning\u003C\u002Fb>\u003C\u002Fi>, Various, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FCVPR-2023-blue\" alt=\"CVPR Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FLAION-AI\u002Fscaling-laws-openclip\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FLAION-AI\u002Fscaling-laws-openclip.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n\n\u003Cb>Video and Temporal Understanding\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Video Understanding with Large Language Models: A Survey\u003C\u002Fb>\u003C\u002Fi>, Various, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.17432\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.12-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fyunlong10\u002FAwesome-LLMs-for-Video-Understanding\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fyunlong10\u002FAwesome-LLMs-for-Video-Understanding.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n### Context Management in Production\n\nIn the agent era, context engineering increasingly means **runtime context management** rather than only prompt construction. Production systems now rely on compaction, caching, artifact-backed state, and scoped instruction loading to keep long-horizon agents efficient and controllable.\n\n\u003Cb>Runtime Context Management Patterns\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>OpenAI Agents Guide\u003C\u002Fb>\u003C\u002Fi>, OpenAI, \u003Ca href=\"https:\u002F\u002Fplatform.openai.com\u002Fdocs\u002Fguides\u002Fagents\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpenAI-2026-blue\" alt=\"OpenAI Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>OpenAI Tools: Conversation State, Prompt Caching, and Compaction\u003C\u002Fb>\u003C\u002Fi>, OpenAI, \u003Ca href=\"https:\u002F\u002Fdevelopers.openai.com\u002Fapi\u002Fdocs\u002Fguides\u002Ftools\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpenAI-2026-blue\" alt=\"OpenAI Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Google ADK: Context Caching and Context Compression\u003C\u002Fb>\u003C\u002Fi>, Google, \u003Ca href=\"https:\u002F\u002Fgoogle.github.io\u002Fadk-docs\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FGoogle-2026-blue\" alt=\"Google Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Claude Code Memory and Scoped Project Instructions\u003C\u002Fb>\u003C\u002Fi>, Anthropic, \u003Ca href=\"https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fdocs\u002Fclaude-code\u002Fmemory\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAnthropic-2026-blue\" alt=\"Anthropic Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LangChain Deep Agents: Filesystem-Based Context Management\u003C\u002Fb>\u003C\u002Fi>, LangChain, \u003Ca href=\"https:\u002F\u002Fdocs.langchain.com\u002Foss\u002Fpython\u002Fdeepagents\u002Foverview\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLangChain-2026-blue\" alt=\"LangChain Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Production Design Questions\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>When should state stay in the prompt versus move into files, memory stores, or external tools?\u003C\u002Fb>\u003C\u002Fi>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>How should long-running threads be compacted without losing provenance, instructions, or active plans?\u003C\u002Fb>\u003C\u002Fi>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>How should project rules be loaded conditionally by path, task, or subagent instead of globally?\u003C\u002Fb>\u003C\u002Fi>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>How should prompt caching be combined with memory writes and retrieval freshness?\u003C\u002Fb>\u003C\u002Fi>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n### Structured Data Integration\n\n\u003Cb>Knowledge Graph-Enhanced Language Models\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Learn Together: Joint Multitask Finetuning of Pretrained KG-enhanced LLM for Downstream Tasks\u003C\u002Fb>\u003C\u002Fi>, Martynova et al., \u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.18653\u002Fv1\u002F2025.genaik-1.2\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICCL-2025.01-blue\" alt=\"ICCL Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FVloods\u002Fmultitask_finetune\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FVloods\u002Fmultitask_finetune.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Knowledge Graph Tuning: Real-time Large Language Model Personalization based on Human Feedback\u003C\u002Fb>\u003C\u002Fi>, Sun et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2025.02-blue\" alt=\"ICLR Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Knowledge Graph-Guided Retrieval Augmented Generation\u003C\u002Fb>\u003C\u002Fi>, Zhu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.06864\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fnju-websoft\u002FKG2RAG\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstarsnju-websoft\u002FKG2RAG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>KGLA: Knowledge Graph Enhanced Language Agents for Customer Service\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.19627\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.10-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\u003Cb>Graph Neural Networks Combined with Language Models\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Are Large Language Models In-Context Graph Learners?\u003C\u002Fb>\u003C\u002Fi>, Li et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.13562\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fyunlong10\u002FAwesome-LLMs-for-Video-Understanding\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fyunlong10\u002FAwesome-LLMs-for-Video-Understanding.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Let's Ask GNN: Empowering Large Language Model for Graph In-Context Learning\u003C\u002Fb>\u003C\u002Fi>, Hu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.07074\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FEMNLP-2024.11-blue\" alt=\"EMNLP Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fppsmk388\u002FAskGNN\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fppsmk388\u002FAskGNN.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>GL-Fusion: Rethinking the Combination of Graph Neural Network and Large Language model\u003C\u002Fb>\u003C\u002Fi>, Yang et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2025.02-blue\" alt=\"ICLR Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>NT-LLM: A Novel Node Tokenizer for Integrating Graph Structure into Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Ji et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.10743\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.10-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\u003Cb>Structured Data Integration\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>CoddLLM: Empowering Large Language Models for Data Analytics\u003C\u002Fb>\u003C\u002Fi>, Authors et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.00329\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Structure-Guided Large Language Models for Text-to-SQL Generation\u003C\u002Fb>\u003C\u002Fi>, Authors et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.13284\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>StructuredRAG: JSON Response Formatting with Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Authors et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2408.11061\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.08-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fweaviate\u002Fstructured-rag\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fweaviate\u002Fstructured-rag.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Foundational KG-LLM Integration Methods\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Unifying Large Language Models and Knowledge Graphs: A Roadmap\u003C\u002Fb>\u003C\u002Fi>, Various, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2306.08302\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.06-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FRManLuo\u002FAwesome-LLM-KG?tab=readme-ov-file\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FRManLuo\u002FAwesome-LLM-KG?tab=readme-ov-file.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Combining Knowledge Graphs and Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Various, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.06564\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.07-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>All Against Some: Efficient Integration of Large Language Models for Message Passing in Graph Neural Networks\u003C\u002Fb>\u003C\u002Fi>, Various, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.14996\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.07-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Large Language Models for Graph Learning\u003C\u002Fb>\u003C\u002Fi>, Various, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FWWW-2024-blue\" alt=\"WWW Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\n### Self-Generated Context\n\n\u003Cb>Self-Supervised Context Generation and Augmentation\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>SelfCite: Self-Supervised Alignment for Context Attribution in Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Chuang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.09604\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Ffacebookresearch\u002FSelfCite\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Ffacebookresearch\u002FSelfCite.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Self-Supervised Prompt Optimization\u003C\u002Fb>\u003C\u002Fi>, Xiang et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FCoRR-2025.01-orange\" alt=\"CoRR Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FFoundationAgents\u002FMetaGPT\u002Ftree\u002Fmain\u002Fexamples\u002Fspo\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FFoundationAgents\u002FMetaGPT\u002Ftree\u002Fmain\u002Fexamples\u002Fspo.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>SCOPE: A Self-supervised Framework for Improving Faithfulness in Conditional Text Generation\u003C\u002Fb>\u003C\u002Fi>, Duong et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2025.01-blue\" alt=\"ICLR Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fsngdng\u002Fscope-faithfulness\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fsngdng\u002Fscope-faithfulness.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Reasoning Models That Generate Their Own Context\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Self-Consistency Improves Chain of Thought Reasoning in Language Models\u003C\u002Fb>\u003C\u002Fi>, Wang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2203.11171\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2023.02-blue\" alt=\"ICLR Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Tree of Thoughts: Deliberate Problem Solving with Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Yao et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2305.10601\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fprinceton-nlp\u002Ftree-of-thought-llm\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fprinceton-nlp\u002Ftree-of-thought-llm.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Rethinking Chain-of-Thought from the Perspective of Self-Training\u003C\u002Fb>\u003C\u002Fi>, Wu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.10827\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.12-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fzongqianwu\u002FST-COT\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fzongqianwu\u002FST-COT.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Autonomous Tree-search Ability of Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Authors et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.10686\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.10-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FZheyuAqaZhang\u002FAutonomous-Tree-search\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FZheyuAqaZhang\u002FAutonomous-Tree-search.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\u003Cb>Iterative Context Refinement and Self-Improvement\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Self-Refine: Iterative Refinement with Self-Feedback\u003C\u002Fb>\u003C\u002Fi>, Madaan et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2303.17651\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.03-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmadaan\u002Fself-refine\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmadaan\u002Fself-refine.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Reflect, Retry, Reward: Self-Improving LLMs via Reinforcement Learning\u003C\u002Fb>\u003C\u002Fi>, Authors et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2505.24726\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Large Language Models Can Self-Improve in Long-context Reasoning\u003C\u002Fb>\u003C\u002Fi>, Li et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2411.08147\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.11-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FSihengLi99\u002FSEALONG\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FSihengLi99\u002FSEALONG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Code Generation with AlphaCodium: From Prompt Engineering to Flow Engineering\u003C\u002Fb>\u003C\u002Fi>, Oren et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.08500\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa> \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FCodium-ai\u002Falphacodium\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FCodium-ai\u002Falphacodium.svg?style=social\" alt=\"GitHub stars\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Language Agent Tree Search Unifies Reasoning Acting and Planning in Language Models\u003C\u002Fb>\u003C\u002Fi>, Zhou et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.04406\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.10-red\" alt=\"arXiv Badge\">\u003C\u002Fa> \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fandyz245\u002FLanguage-Agent-Tree-Search\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fandyz245\u002FLanguage-Agent-Tree-Search.svg?style=social\" alt=\"GitHub stars\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\u003Cb>Meta-Learning and Autonomous Context Evolution\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Meta-in-context learning in large language models\u003C\u002Fb>\u003C\u002Fi>, Coda-Forno et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNeurIPS-2023.12-blue\" alt=\"NeurIPS Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>EvoPrompt: Connecting LLMs with Evolutionary Algorithms Yields Powerful Prompt Optimizers\u003C\u002Fb>\u003C\u002Fi>, Guo et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2309.08532\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2024.01-blue\" alt=\"ICLR Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fbeeevita\u002FEvoPrompt\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fbeeevita\u002FEvoPrompt.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>AutoPDL: Automatic Prompt Optimization for LLM Agents\u003C\u002Fb>\u003C\u002Fi>, Spiess et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2504.04365\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAutoML-2025.04-orange\" alt=\"AutoML Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Agent-Pro: Learning to Evolve Coder Agents via Proposal-based Programming\u003C\u002Fb>\u003C\u002Fi>, Zhang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.17574\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\u003Cb>Foundational Chain-of-Thought Research\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Chain-of-thought prompting elicits reasoning in large language models\u003C\u002Fb>\u003C\u002Fi>, Wei et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNeurIPS-2022-blue\" alt=\"NeurIPS Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\n---\n\n## 🛠️ Implementation and Challenges\n\n### 0. Agent Harnesses and Runtime Systems\n\nIn 2026, many of the most important advances in context engineering no longer live only inside the prompt. They live inside the **agent harness**: the runtime loop that manages plans, subagents, checkpoints, files, approvals, tool execution, and recovery from failure. This is where context engineering becomes agent engineering.\n\n\u003Cb>Harness and Runtime Design References\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Building Effective Agents\u003C\u002Fb>\u003C\u002Fi>, Anthropic, \u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Fbuilding-effective-agents\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAnthropic-2024.12-blue\" alt=\"Anthropic Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>OpenAI Agents Guide\u003C\u002Fb>\u003C\u002Fi>, OpenAI, \u003Ca href=\"https:\u002F\u002Fplatform.openai.com\u002Fdocs\u002Fguides\u002Fagents\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpenAI-2026-blue\" alt=\"OpenAI Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Google Agent Development Kit (ADK)\u003C\u002Fb>\u003C\u002Fi>, Google, \u003Ca href=\"https:\u002F\u002Fgoogle.github.io\u002Fadk-docs\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FGoogle-2026-blue\" alt=\"Google Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LangChain Deep Agents Overview\u003C\u002Fb>\u003C\u002Fi>, LangChain, \u003Ca href=\"https:\u002F\u002Fdocs.langchain.com\u002Foss\u002Fpython\u002Fdeepagents\u002Foverview\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLangChain-2026-blue\" alt=\"LangChain Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Microsoft Agent Framework Overview\u003C\u002Fb>\u003C\u002Fi>, Microsoft, \u003Ca href=\"https:\u002F\u002Flearn.microsoft.com\u002Fen-us\u002Fagent-framework\u002Fuser-guide\u002Foverview\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FMicrosoft-2026-blue\" alt=\"Microsoft Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Core Runtime Concerns\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Planning and decomposition\u003C\u002Fb>\u003C\u002Fi>: how long tasks are split into manageable units\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Durable execution\u003C\u002Fb>\u003C\u002Fi>: how agent state is checkpointed, resumed, or replayed\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Context isolation\u003C\u002Fb>\u003C\u002Fi>: how subagents and tools avoid polluting each other's working state\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Sandboxing and artifacts\u003C\u002Fb>\u003C\u002Fi>: how file systems, shells, browsers, and outputs become part of the context pipeline\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Human approvals and interrupts\u003C\u002Fb>\u003C\u002Fi>: how production agents remain controllable during risky or long-running actions\u003C\u002Fli>\n\u003C\u002Ful>\n\n### 1. Retrieval-Augmented Generation (RAG)\n\n\u003Cb>survey\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Retrieval-Augmented Generation for Large Language Models: A Survey\u003C\u002Fb>\u003C\u002Fi>, Yunfan Gao et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.10997\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.12-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FTongji-KGLLM\u002FRAG-Survey\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FTongji-KGLLM\u002FRAG-Survey.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>A Survey of Graph Retrieval-Augmented Generation for Customized Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Siyun Zhao et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.13958\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.01-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FDEEP-PolyU\u002FAwesome-GraphRAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FDEEP-PolyU\u002FAwesome-GraphRAG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Retrieval Augmented Generation (RAG) and Beyond: A Comprehensive Survey on How to Make your LLMs use External Data More Wisely\u003C\u002Fb>\u003C\u002Fi>, Siyun Zhao et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2409.14924\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.09-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Evaluation of Retrieval-Augmented Generation: A Survey\u003C\u002Fb>\u003C\u002Fi>, Hao Yu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.07437\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.07-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FYHPeter\u002FAwesome-RAG-Evaluation\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FYHPeter\u002FAwesome-RAG-Evaluation.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks\u003C\u002Fb>\u003C\u002Fi>, Lewis et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2005.11401\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2020.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fcostadev00\u002FRAG-paper-implementation-from-scratch\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fcostadev00\u002FRAG-paper-implementation-from-scratch.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>A Survey on Knowledge-Oriented Retrieval-Augmented Generation\u003C\u002Fb>\u003C\u002Fi>, Cheng et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2503.10677\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FUSTCAGI\u002FAwesome-Papers-Retrieval-Augmented-Generation\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FUSTCAGI\u002FAwesome-Papers-Retrieval-Augmented-Generation.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>A Survey on RAG Meeting LLMs: Towards Retrieval-Augmented Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Ding et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.06211\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.06-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n\u003Cb>Naive RAG\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Beyond the Limits: A Survey of Techniques to Extend the Context Length in Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Xindi Wang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.02244\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>In-context Examples Selection for Machine Translation\u003C\u002Fb>\u003C\u002Fi>, Sweta Agrawal et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2212.02437\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>In Defense of RAG in the Era of Long-Context Language Models\u003C\u002Fb>\u003C\u002Fi>, Tan Yu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2409.01666\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.09-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks\u003C\u002Fb>\u003C\u002Fi>, Patrick Lewis et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2005.11401\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2020.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LightRAG: Simple and Fast Retrieval-Augmented Generation\u003C\u002Fb>\u003C\u002Fi>, Zirui Guo et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.05779\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fanonymous.4open.science\u002Fr\u002FLightRAG-2BEE\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fanonymous\u002FLightRAG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Generate rather than Retrieve: Large Language Models are Strong Context Generators\u003C\u002Fb>\u003C\u002Fi>, Wenhao Yu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2209.10063\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.09-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fwyu97\u002FGenRead\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fwyu97\u002FGenRead.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Large language models can be easily distracted by irrelevant context\u003C\u002Fb>\u003C\u002Fi>, Freda Shi et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2302.00093\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fgoogle-research-datasets\u002FGSM-IC\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fgoogle-research-datasets\u002FGSM-IC.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Old IR Methods Meet RAG\u003C\u002Fb>\u003C\u002Fi>, Oz Huly et al.\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Dense Passage Retrieval for Open-Domain Question Answering\u003C\u002Fb>\u003C\u002Fi>, Vladimir Karpukhin et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2004.04906\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2020.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Ffacebookresearch\u002FDPR\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Ffacebookresearch\u002FDPR.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n\u003Cb>Advanced RAG\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Adaptive-RAG: Learning to Adapt Retrieval-Augmented Large Language Models through Question Complexity\u003C\u002Fb>\u003C\u002Fi>, Soyeong Jeong et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2403.14403\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fstarsuzi\u002FAdaptive-RAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fstarsuzi\u002FAdaptive-RAG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Improving language models by retrieving from trillions of tokens\u003C\u002Fb>\u003C\u002Fi>, Sebastian Borgeaud et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2112.04426\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>FoRAG: Factuality-optimized Retrieval Augmented Generation for Web-enhanced Long-form Question Answering\u003C\u002Fb>\u003C\u002Fi>, Tianchi Cai et al.\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>IM-RAG: Multi-Round Retrieval-Augmented Generation Through Learning Inner Monologues\u003C\u002Fb>\u003C\u002Fi>, Diji Yang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.13021\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>RAGCache: Efficient Knowledge Caching for Retrieval-Augmented Generation\u003C\u002Fb>\u003C\u002Fi>, Chao Jin et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.12457\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Corrective Retrieval Augmented Generation\u003C\u002Fb>\u003C\u002Fi>, Shi-Qi Yan et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.15884\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FHuskyInSalt\u002FCRAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FHuskyInSalt\u002FCRAG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>RankRAG: Unifying Context Ranking with Retrieval-Augmented Generation in LLMs\u003C\u002Fb>\u003C\u002Fi>, Yue Yu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.02485\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Astute RAG: Overcoming Imperfect Retrieval Augmentation and Knowledge Conflicts for Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Fei Wang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.07176\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Learning to Filter Context for Retrieval-Augmented Generation\u003C\u002Fb>\u003C\u002Fi>, Zhiruo Wang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2311.08377\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fzorazrw\u002Ffilco\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fzorazrw\u002Ffilco.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Query Rewriting in Retrieval-Augmented Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Xinbei Ma et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2305.14283\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fqijimrc\u002FROBUST\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fqijimrc\u002FROBUST.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>UPRISE: Universal Prompt Retrieval for Improving Zero-Shot Evaluation\u003C\u002Fb>\u003C\u002Fi>, Daixuan Cheng et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2303.08518\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FMatthewKKai\u002FSMRC2\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FMatthewKKai\u002FSMRC2.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Longllmlingua: Accelerating and enhancing llms in long context scenarios via prompt compression\u003C\u002Fb>\u003C\u002Fi>, Huiqiang Jiang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.06839\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002FLLMLingua\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmicrosoft\u002FLLMLingua.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Document-level event argument extraction by conditional generation\u003C\u002Fb>\u003C\u002Fi>, Sha Li et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2104.05919\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2021.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fraspberryice\u002Fgen-arg\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fraspberryice\u002Fgen-arg.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Multi-sentence Argument Linking\u003C\u002Fb>\u003C\u002Fi>, Seth Ebner et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F1911.03766\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2019.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fnlp.jhu.edu\u002Frams\u002F\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fnlp-jhu\u002FRAMS.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Fine-tuning or retrieval? comparing knowledge injection in llms\u003C\u002Fb>\u003C\u002Fi>, Oded Ovadia et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.05934\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>IAG: Induction-Augmented Generation Framework for Answering Reasoning Questions\u003C\u002Fb>\u003C\u002Fi>, Zhebin Zhang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2311.18397\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Retrieval Meets Long Context Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Peng Xu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.03025\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Dense x retrieval: What retrieval granularity should we use?\u003C\u002Fb>\u003C\u002Fi>, Tong Chen et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.06648\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fct123098\u002Ffactoid-wiki\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fct123098\u002Ffactoid-wiki.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Investigating the Factual Knowledge Boundary of Large Language Models with Retrieval Augmentation\u003C\u002Fb>\u003C\u002Fi>, Ruiyang Ren et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2307.11019\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FRUCAIBox\u002FLLM-Knowledge-Boundary\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FRUCAIBox\u002FLLM-Knowledge-Boundary.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>The Power of Noise: Redefining Retrieval for RAG Systems\u003C\u002Fb>\u003C\u002Fi>, Florin Cuconasu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.14887\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fflorin-git\u002FThe-Power-of-Noise\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fflorin-git\u002FThe-Power-of-Noise.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>RECITATION-AUGMENTED LANGUAGE MODELS\u003C\u002Fb>\u003C\u002Fi>, Zhiqing Sun et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.01296\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FEdward-Sun\u002FRECITE\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FEdward-Sun\u002FRECITE.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Robust Retrieval Augmented Generation for Zero-shot Slot Filling\u003C\u002Fb>\u003C\u002Fi>, Michael Glass et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2108.13934\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2021.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FIBM\u002Fkgi-slot-filling\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FIBM\u002Fkgi-slot-filling.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>In-Context Retrieval-Augmented Language Models\u003C\u002Fb>\u003C\u002Fi>, Ori Ram et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2302.00083\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FAI21Labs\u002Fin-context-ralm\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FAI21Labs\u002Fin-context-ralm.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Learning to Retrieve In-Context Examples for Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Liang Wang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2307.07164\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.03-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002FLMOps\u002Ftree\u002Fmain\u002Fllm_retriever\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmicrosoft\u002FLMOps.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n\u003Cb>Modular RAG\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>FlashRAG: A Modular Toolkit for Efficient Retrieval-Augmented Generation Research\u003C\u002Fb>\u003C\u002Fi>, Jiajie Jin et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.13576\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FRUC-NLPIR\u002FFlashRAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FRUC-NLPIR\u002FFlashRAG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Multi-Head RAG: Solving Multi-Aspect Problems with LLMs\u003C\u002Fb>\u003C\u002Fi>, Maciej Besta et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2406.05085\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fspcl\u002FMRAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fspcl\u002FMRAG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>StructRAG: Boosting Knowledge Intensive Reasoning of LLMs via Inference-time Hybrid Information Structurization\u003C\u002Fb>\u003C\u002Fi>, Zhuoqun Li et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.08815\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FLi-Z-Q\u002FStructRAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FLi-Z-Q\u002FStructRAG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>RAFT: Adapting Language Model to Domain Specific RAG\u003C\u002Fb>\u003C\u002Fi>, Tianjun Zhang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2403.10131\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FShishirPatil\u002Fgorilla\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FShishirPatil\u002Fgorilla.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Retrieval-Generation Alignment for End-to-End Task-Oriented Dialogue System\u003C\u002Fb>\u003C\u002Fi>, Weizhou Shen et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.08877\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fshenwzh3\u002FMK-TOD\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fshenwzh3\u002FMK-TOD.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>UniMS-RAG: A Unified Multi-source Retrieval-Augmented Generation for Personalized Dialogue Systems\u003C\u002Fb>\u003C\u002Fi>, Hongru Wang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.13256\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Retrieve-and-Sample: Document-level Event Argument Extraction via Hybrid Retrieval Augmentation\u003C\u002Fb>\u003C\u002Fi>, Yubing Ren et al.\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>RA-DIT: RETRIEVAL-AUGMENTED DUAL INSTRUCTION TUNING\u003C\u002Fb>\u003C\u002Fi>, Xi Victoria Lin et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.01352\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Ffacebookresearch\u002FRA-DIT\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Ffacebookresearch\u002FRA-DIT.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Self-Knowledge Guided Retrieval Augmentation for Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Yile Wang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.05002\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FTHUNLP-MT\u002FSKR\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FTHUNLP-MT\u002FSKR.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Prompt-Guided Retrieval Augmentation for Non-Knowledge-Intensive Tasks\u003C\u002Fb>\u003C\u002Fi>, Zhicheng Guo et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2305.17653\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FTHUNLP-MT\u002FPGRA\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FTHUNLP-MT\u002FPGRA.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>REPLUG: Retrieval-Augmented Black-Box Language Models\u003C\u002Fb>\u003C\u002Fi>, Weijia Shi et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2301.12652\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Query Rewriting for Retrieval-Augmented Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Xinbei Ma et al., \u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.18653\u002Fv1\u002F2023.emnlp-main.323\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FEMNLP-2023.00-blue\" alt=\"DOI Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fxbmxb\u002FRAG-query-rewriting\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fxbmxb\u002FRAG-query-rewriting.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Lift Yourself Up: Retrieval-augmented Text Generation with Self-Memory\u003C\u002Fb>\u003C\u002Fi>, Xin Cheng et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2305.02437\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FHannibal046\u002FSelfMemory\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FHannibal046\u002FSelfMemory.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Improving the Domain Adaptation of Retrieval Augmented Generation (RAG) Models for Open Domain Question Answering\u003C\u002Fb>\u003C\u002Fi>, Shamane Siriwardhana et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.02627\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n\u003Cb>Graph-Based RAG\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Don't Forget to Connect! Improving RAG with Graph-based Reranking\u003C\u002Fb>\u003C\u002Fi>, Jialin Dong et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.18414\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>From Local to Global: A Graph RAG Approach to Query-Focused Summarization\u003C\u002Fb>\u003C\u002Fi>, Darren Edge et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.16130\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>GRAG: Graph Retrieval-Augmented Generation\u003C\u002Fb>\u003C\u002Fi>, Yuntong Hu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.16506\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FHuieL\u002FGRAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FHuieL\u002FGRAG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Iseeq: Information seeking question generation using dynamic meta-information retrieval and knowledge graphs\u003C\u002Fb>\u003C\u002Fi>, Manas Gaur et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2112.07622\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.06-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmanasgaur\u002FAAAI-22\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmanasgaur\u002FAAAI-22.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>G-retriever: Retrieval-augmented generation for textual graph understanding and question answering\u003C\u002Fb>\u003C\u002Fi>, Xiaoxin He et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.07630\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FXiaoxinHe\u002FG-Retriever\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FXiaoxinHe\u002FG-Retriever.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Knowledge graph prompting for multi-document question answering\u003C\u002Fb>\u003C\u002Fi>, Yu Wang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.08774\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FYuWVandy\u002FKG-LLM-MDQA\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FYuWVandy\u002FKG-LLM-MDQA.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>GNN-RAG: Graph Neural Retrieval for Large Language Model Reasoning\u003C\u002Fb>\u003C\u002Fi>, Costas Mavromatis et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.20139\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fcmavro\u002FGNN-RAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fcmavro\u002FGNN-RAG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LightPROF: A Lightweight Reasoning Framework for Large Language Model on Knowledge Graph\u003C\u002Fb>\u003C\u002Fi>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Ftsinghua-fib-lab\u002FACL24-EconAgent\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Ftsinghua-fib-lab\u002FACL24-EconAgent.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Simple Is Effective: The Roles of Graphs and Large Language Models in Knowledge-Graph-Based Retrieval-Augmented Generation\u003C\u002Fb>\u003C\u002Fi>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FGraph-COM\u002FSubgraphRAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FGraph-COM\u002FSubgraphRAG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Knowledge Graph-Guided Retrieval Augmented Generation\u003C\u002Fb>\u003C\u002Fi>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fnju-websoft\u002FKG2RAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fnju-websoft\u002FKG2RAG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>MedRAG: Enhancing Retrieval-augmented Generation with Knowledge Graph-Elicited Reasoning for Healthcare Copilot\u003C\u002Fb>\u003C\u002Fi>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FSNOWTEAM2023\u002FMedRAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FSNOWTEAM2023\u002FMedRAG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Mitigating Large Language Model Hallucinations via Autonomous Knowledge Graph-based Retrofitting\u003C\u002Fb>\u003C\u002Fi>, KGR et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2311.13314\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.11-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmansicer\u002FMAIC\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmansicer\u002FMAIC.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>In-depth Analysis of Graph-based RAG in a Unified Framework\u003C\u002Fb>\u003C\u002Fi>\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2503.04338\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FJayLZhou\u002FGraphRAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FJayLZhou\u002FGraphRAG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval\u003C\u002Fb>\u003C\u002Fi>, Parth Sarthi et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.18059\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fparthsarthi03\u002Fraptor\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fparthsarthi03\u002Fraptor.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>TableRAG: Million-Token Table Understanding with Language Models\u003C\u002Fb>\u003C\u002Fi>, Si-An Chen et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.04739\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fgoogle-research\u002Fgoogle-research\u002Ftree\u002Fmaster\u002Ftable_rag\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fgoogle-research\u002Fgoogle-research.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>KAG: Boosting LLMs in Professional Domains via Knowledge Augmented Generation\u003C\u002Fb>\u003C\u002Fi>, Lei Liang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2409.13731\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FOpenSPG\u002FKAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FOpenSPG\u002FKAG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>GFM-RAG: Graph Foundation Model for Retrieval Augmented Generation\u003C\u002Fb>\u003C\u002Fi>, Luo et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.01113\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FRManLuo\u002Fgfm-rag\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FRManLuo\u002Fgfm-rag.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>HybridRAG: A Hybrid Retrieval System for RAG Combining Vector and Graph Search\u003C\u002Fb>\u003C\u002Fi>, Sarabesh, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FGitHub-2024.12-white\" alt=\"GitHub Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fsarabesh\u002FHybridRAG\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fsarabesh\u002FHybridRAG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n\u003Cb>Agentic RAG\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>From RAG to Memory: Non-Parametric Continual Learning for Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Bernal Jiménez Gutiérrez et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.14802\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FOSU-NLP-Group\u002FHippoRAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FOSU-NLP-Group\u002FHippoRAG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Bernal Jiménez Gutiérrez et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.14924\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FOSU-NLP-Group\u002FHippoRAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FOSU-NLP-Group\u002FHippoRAG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>GraphReader: Building Graph-based Agent to Enhance Long-Context Abilities of Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Shilong Li et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2406.14550\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>PlanRAG: A Plan-then-Retrieval Augmented Generation for Generative Large Language Models as Decision Makers\u003C\u002Fb>\u003C\u002Fi>, Myeonghwa Lee et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2406.12430\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmyeon9h\u002FPlanRAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmyeon9h\u002FPlanRAG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection\u003C\u002Fb>\u003C\u002Fi>, Akari Asai et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.08353\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FAkariAsai\u002Fself-rag\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FAkariAsai\u002Fself-rag.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>DeepRAG: Thinking to Retrieve Step by Step for Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Xinyan Guan et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.01142\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Paperqa: Retrieval-augmented generative agent for scientific research\u003C\u002Fb>\u003C\u002Fi>, Jakub Lála et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.07559\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Large Language Models as Source Planner for Personalized Knowledge-grounded Dialogues\u003C\u002Fb>\u003C\u002Fi>, Hongru Wang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2308.06181\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fhrwise-nlp\u002FSAFARI\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fhrwise-nlp\u002FSAFARI.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>PRCA: Fitting Black-Box Large Language Models for Retrieval Question Answering via Pluggable Reward-Driven Contextual Adapter\u003C\u002Fb>\u003C\u002Fi>, Haoyan Yang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.18347\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fxbmxb\u002FRAG-query-rewriting\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fxbmxb\u002FRAG-query-rewriting.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>SELF-RAG: LEARNING TO RETRIEVE, GENERATE, AND CRITIQUE THROUGH SELF-REFLECTION\u003C\u002Fb>\u003C\u002Fi>, Akari Asai et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.11511\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fselfrag.github.io\u002F\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fselfrag\u002Fselfrag.github.io.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>RAT: Retrieval Augmented Thoughts Elicit Context-Aware Reasoning in Long-Horizon Generation\u003C\u002Fb>\u003C\u002Fi>, Zihao Wang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2403.05313\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FCraftJarvis\u002FRAT\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FCraftJarvis\u002FRAT.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Chain-of-verification reduces hallucination in large language models\u003C\u002Fb>\u003C\u002Fi>, Shehzaad Dhuliawala et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2309.11495\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>HM-RAG: Hierarchical Multi-Agent Multimodal Retrieval Augmented Generation\u003C\u002Fb>\u003C\u002Fi>, Liu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2504.12330\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Focean-luna\u002FHMRAG\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Focean-luna\u002FHMRAG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>MultiHop-RAG: Benchmarking Retrieval-Augmented Generation for Multi-Hop Queries\u003C\u002Fb>\u003C\u002Fi>, Tang & Yang, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.15391\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.01-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fyixuantt\u002FMultiHop-RAG\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fyixuantt\u002FMultiHop-RAG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>MMOA-RAG: Improving Retrieval-Augmented Generation through Multi-Agent Reinforcement Learning\u003C\u002Fb>\u003C\u002Fi>, Chen et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2010.10110\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2021.01-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fchenyiqun\u002FMMOA-RAG\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fchenyiqun\u002FMMOA-RAG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Search-in-the-Chain: Towards Accurate, Credible, and Up-to-Date Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Menick et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2304.14732\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n\u003Cb>Real-Time and Streaming RAG\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>StreamingRAG: Real-time Contextual Retrieval and Generation Framework\u003C\u002Fb>\u003C\u002Fi>, Sankaradas et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.14101\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.01-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fvideo-db\u002FStreamRAG\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fvideo-db\u002FStreamRAG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Multi-task Retriever Fine-tuning for Domain-Specific and Efficient RAG\u003C\u002Fb>\u003C\u002Fi>, Authors, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.04652\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.01-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\n### 2. Memory Systems\n\n#### Runtime Memory Design Patterns\n\nModern memory systems are no longer a single retrieval store. Production agents increasingly separate:\n\n- **Session \u002F thread state** for active work in progress\n- **Long-term semantic memory** for user or project facts\n- **Episodic memory** for trajectories, past actions, and reusable experiences\n- **Procedural memory** for learned workflows, instructions, and stable operating preferences\n\n\u003Cb>Memory Design References\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>LangGraph Memory Overview\u003C\u002Fb>\u003C\u002Fi>, LangChain, \u003Ca href=\"https:\u002F\u002Fdocs.langchain.com\u002Foss\u002Fjavascript\u002Flanggraph\u002Fmemory\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLangChain-2026-blue\" alt=\"LangChain Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Letta Memory Blocks\u003C\u002Fb>\u003C\u002Fi>, Letta, \u003Ca href=\"https:\u002F\u002Fdocs.letta.com\u002Fguides\u002Fcore-concepts\u002Fmemory\u002Fmemory-blocks\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLetta-2026-blue\" alt=\"Letta Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Claude Code Memory\u003C\u002Fb>\u003C\u002Fi>, Anthropic, \u003Ca href=\"https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fdocs\u002Fclaude-code\u002Fmemory\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAnthropic-2026-blue\" alt=\"Anthropic Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n#### Project Memory and Instruction Artifacts\n\nCoding agents have made project memory concrete. In practice, memory now often lives in artifacts such as repository instruction files, scoped rules, reusable skills, and long-lived project notes rather than only in vector stores.\n\n\u003Cb>Project Memory References\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Introducing Codex\u003C\u002Fb>\u003C\u002Fi>, OpenAI, \u003Ca href=\"https:\u002F\u002Fopenai.com\u002Findex\u002Fintroducing-codex\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpenAI-2025.05-blue\" alt=\"OpenAI Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Claude Code Memory\u003C\u002Fb>\u003C\u002Fi>, Anthropic, \u003Ca href=\"https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fdocs\u002Fclaude-code\u002Fmemory\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAnthropic-2026-blue\" alt=\"Anthropic Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Claude Code Subagents\u003C\u002Fb>\u003C\u002Fi>, Anthropic, \u003Ca href=\"https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fdocs\u002Fclaude-code\u002Fsub-agents\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAnthropic-2026-blue\" alt=\"Anthropic Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LangChain Deep Agents Overview\u003C\u002Fb>\u003C\u002Fi>, LangChain, \u003Ca href=\"https:\u002F\u002Fdocs.langchain.com\u002Foss\u002Fpython\u002Fdeepagents\u002Foverview\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLangChain-2026-blue\" alt=\"LangChain Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Persistent Memory Architecture\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>MemGPT: Towards LLMs as Operating Systems\u003C\u002Fb>\u003C\u002Fi>, Packer et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.08560\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.10-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fletta-ai\u002Fletta\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fletta-ai\u002Fletta.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory\u003C\u002Fb>\u003C\u002Fi>, Taranjeet et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2504.19413\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmem0ai\u002Fmem0\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmem0ai\u002Fmem0.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>MemoryLLM: Towards Self-Updatable Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Wang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.04624\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fwangyu-ustc\u002FMemoryLLM\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fwangyu-ustc\u002FMemoryLLM.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Infinite-LLM: Efficient LLM Service for Long Context with DistAttention and Distributed KVCache\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.02669\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.01-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Memory-Augmented Generative Adversarial Transformers\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.19218\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Memory Interchange Standards\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>PAM (Portable AI Memory): An Open Interchange Format for AI User Memories\u003C\u002Fb>\u003C\u002Fi>, Daniel Gines, \u003Ca href=\"https:\u002F\u002Fportable-ai-memory.org\u002Fspec\u002Fv1.0\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FSpec-v1.0-blue\" alt=\"Spec Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fportable-ai-memory\u002Fpython-sdk\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fportable-ai-memory\u002Fpython-sdk.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Memory-Augmented Neural Networks\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Survey on Memory-Augmented Neural Networks: Cognitive Insights to AI Applications\u003C\u002Fb>\u003C\u002Fi>, Khosla et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.06141\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.12-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>A Machine with Short-Term, Episodic, and Semantic Memory Systems\u003C\u002Fb>\u003C\u002Fi>, Kim et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2212.02098\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.12-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fhumemai\u002Fagent-room-env-v1\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fhumemai\u002Fagent-room-env-v1.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>From Human Memory to AI Memory: A Survey on Memory Mechanisms in the Era of LLMs\u003C\u002Fb>\u003C\u002Fi>, Wu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2504.15965\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\u003Cb>Episodic Memory and Context Persistence\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>The Role of Memory in LLMs: Persistent Context for Smarter Conversations\u003C\u002Fb>\u003C\u002Fi>, Porcu, \u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.18535\u002Fijsrm\u002Fv12i11.ec04\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FIJSRM-2024.11-blue\" alt=\"IJSRM Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Episodic Memory in AI Agents Poses Risks that Should Be Studied and Mitigated\u003C\u002Fb>\u003C\u002Fi>, Christiano et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.11739\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.10-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Larimar: Large Language Models with Episodic Memory Control\u003C\u002Fb>\u003C\u002Fi>, Goyal et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2403.11901\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICML-2024.03-blue\" alt=\"ICML Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>EM-LLM: Human-like Episodic Memory for Infinite Context LLMs\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.09450\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2024.07-blue\" alt=\"ICLR Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fem-llm\u002FEM-LLM-model\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fem-llm\u002FEM-LLM-model.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Large Language Models with Controllable Working Memory\u003C\u002Fb>\u003C\u002Fi>, Goyal et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2211.05110\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.11-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Empowering Working Memory for Large Language Model Agents\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.17259\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.12-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Continual Learning and Memory Consolidation\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Prediction Error-Driven Memory Consolidation for Continual Learning\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNeurIPS-2020.11-blue\" alt=\"NeurIPS Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Overcoming Catastrophic Forgetting in Continual Learning by Exploring Eigenvalues of Hessian Matrix\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNeurIPS-2023-blue\" alt=\"NeurIPS Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Probabilistic Metaplasticity for Continual Learning with Memristors in Spiking Networks\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.11-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Conversational Memory\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>MemoChat: Tuning LLMs to Use Memos for Consistent Long-Range Open-Domain Conversation\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2308.08239\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.08-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Think-in-Memory: Recalling and Post-thinking Enable LLMs with Long-Term Memory\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2311.08719\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.11-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Generative Agents: Interactive Simulacra of Human Behavior\u003C\u002Fb>\u003C\u002Fi>, Park et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2304.03442\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Self-Controlled Memory Framework for Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2304.13343\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Personalization and Memory\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Personalized LLM Response Generation with Parameterized User Memory Injection\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.03565\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Soul-Driven Interaction Design: A Position Paper on Declarative Persona Specifications for AI Agents\u003C\u002Fb>\u003C\u002Fi>, Lee, \u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.5281\u002Fzenodo.18678616\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FZenodo-2026.02-blue\" alt=\"Zenodo Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Soul Spec — Open Specification for AI Agent Persona Packages\u003C\u002Fb>\u003C\u002Fi>, ClawSouls, \u003Ca href=\"https:\u002F\u002Fclawsouls.ai\u002Fspec\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FSpec-v0.4-blue\" alt=\"Spec Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fclawsouls\u002Fsoul-spec-mcp\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fclawsouls\u002Fsoul-spec-mcp.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Safety and Alignment with Memory\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Constitutional AI: Harmlessness from AI Feedback\u003C\u002Fb>\u003C\u002Fi>, Bai et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2212.08073\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.12-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Improving alignment of dialogue agents via targeted human judgements (Sparrow)\u003C\u002Fb>\u003C\u002Fi>, Glaese et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2209.14375\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.09-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Tool Integration and Memory\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>WebGPT: Browser-assisted question-answering with human feedback\u003C\u002Fb>\u003C\u002Fi>, Nakano et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2112.09332\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2021.12-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs\u003C\u002Fb>\u003C\u002Fi>, Qin et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2307.16789\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.07-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Learning and Reflection\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Language Models are Few-Shot Learners (GPT-3)\u003C\u002Fb>\u003C\u002Fi>, Brown et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2005.14165\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2020.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Reflexion: Language Agents with Verbal Reinforcement Learning\u003C\u002Fb>\u003C\u002Fi>, Shinn et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2303.11366\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNeurIPS-2023.03-blue\" alt=\"NeurIPS Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fnoahshinn\u002Freflexion\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fnoahshinn\u002Freflexion.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n### 3. Agent Communication\n\n\u003Cb>Survey\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>A Survey of AI Agent Protocols\u003C\u002Fb>\u003C\u002Fi>, Yingxuan Yang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2504.16736\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.06-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fzoe-yyx\u002FAwesome-AIAgent-Protocol\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fzoe-yyx\u002FAwesome-AIAgent-Protocol.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>A Survey of Multi-Agent Deep Reinforcement Learning with Communication\u003C\u002Fb>\u003C\u002Fi>, Changxi Zhu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2203.08975\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.10-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Beyond Self-Talk: A Communication-Centric Survey of LLM-Based Multi-Agent Systems\u003C\u002Fb>\u003C\u002Fi>, Bingyu Yan et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.14321\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.06-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Large Language Model based Multi-Agents: A Survey of Progress and Challenges\u003C\u002Fb>\u003C\u002Fi>, Taicheng Guo et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.01680\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Ftaichengguo\u002FLLM_MultiAgents_Survey_Papers\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Ftaichengguo\u002FLLM_MultiAgents_Survey_Papers.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n#### Open Agent Protocols and Interoperability\n\nOpen protocols have become a major part of agent engineering. In practice, modern agent systems increasingly separate:\n\n- **agent-to-tool protocols** such as MCP\n- **agent-to-agent protocols** such as A2A and ACP-style remote invocation\n- **agent-to-UI protocols** such as AG-UI\n- **portable agent definitions** such as AgentSchema\n\n\u003Cb>Official Protocol and Interoperability References\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Model Context Protocol Specification\u003C\u002Fb>\u003C\u002Fi>, MCP Working Group, \u003Ca href=\"https:\u002F\u002Fmodelcontextprotocol.io\u002Fspecification\u002F2025-06-18\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FSpec-2025.06-blue\" alt=\"Spec Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Model Context Protocol Architecture\u003C\u002Fb>\u003C\u002Fi>, MCP Working Group, \u003Ca href=\"https:\u002F\u002Fmodelcontextprotocol.io\u002Fdocs\u002Flearn\u002Farchitecture\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FDocs-2026-blue\" alt=\"Docs Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Agent2Agent Protocol (A2A)\u003C\u002Fb>\u003C\u002Fi>, Google, \u003Ca href=\"https:\u002F\u002Fa2a-protocol.org\u002Flatest\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FProtocol-2026-blue\" alt=\"Protocol Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>AG-UI Documentation\u003C\u002Fb>\u003C\u002Fi>, CopilotKit Team, \u003Ca href=\"https:\u002F\u002Fdocs.ag-ui.com\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FProtocol-2026-blue\" alt=\"Protocol Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>ACP Connect\u003C\u002Fb>\u003C\u002Fi>, AGNTCY, \u003Ca href=\"https:\u002F\u002Fdocs.agntcy.org\u002Fsyntactic\u002Fconnect\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FProtocol-2026-blue\" alt=\"Protocol Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>AgentSchema\u003C\u002Fb>\u003C\u002Fi>, Microsoft, \u003Ca href=\"https:\u002F\u002Fmicrosoft.github.io\u002FAgentSchema\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FSchema-2026-blue\" alt=\"Schema Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Agent Interoperability Protocols\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>A survey of agent interoperability protocols: Model Context Protocol (MCP), Agent Communication Protocol (ACP), and Agent-to-Agent Protocol (A2A)\u003C\u002Fb>\u003C\u002Fi>, Zhang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2505.02279\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Expressive Multi-Agent Communication via Identity-Aware Learning\u003C\u002Fb>\u003C\u002Fi>, Du et al., \u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.1609\u002Faaai.v38i16.29683\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAAAI-2024.03-blue\" alt=\"AAAI Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Context-aware Communication for Multi-agent Reinforcement Learning (CACOM)\u003C\u002Fb>\u003C\u002Fi>, Li et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.15600\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.12-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FLXXXXR\u002FCACOM\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FLXXXXR\u002FCACOM.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>A survey of agent interoperability protocols: Model Context Protocol (MCP), Agent Communication Protocol (ACP), Agent-to-Agent Protocol (A2A), and Agent Network Protocol (ANP)\u003C\u002Fb>\u003C\u002Fi>, Abul Ehtesham et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2505.02279\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Agent Capability Negotiation and Binding Protocol (ACNBP)\u003C\u002Fb>\u003C\u002Fi>, Ken Huang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2506.13590\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.06-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>A Scalable Communication Protocol for Networks of Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Samuele Marro et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.11905\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.10-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fagora-protocol\u002Fpaper-demo\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fagora-protocol\u002Fpaper-demo.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Model Context Protocol (MCP)\u003C\u002Fb>\u003C\u002Fi>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmodelcontextprotocol\u002Fmodelcontextprotocol\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmodelcontextprotocol\u002Fmodelcontextprotocol.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Agent2Agent (A2A) Protocol\u003C\u002Fb>\u003C\u002Fi>\n    \u003Ca href=\"https:\u002F\u002Fgoogle.github.io\u002FA2A\u002F\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fgoogle\u002FA2A.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Agent Network Protocol (ANP)\u003C\u002Fb>\u003C\u002Fi>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fagent-network-protocol\u002FAgentNetworkProtocol\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fagent-network-protocol\u002FAgentNetworkProtocol.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n\u003Cb>Structured Communication Frameworks\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Learning Structured Communication for Multi-Agent Reinforcement Learning\u003C\u002Fb>\u003C\u002Fi>, Wang et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAAMAS-2023.05-blue\" alt=\"AAMAS Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fbellmanequation\u002FLSC\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fbellmanequation\u002FLSC.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>AC2C: Adaptively Controlled Two-Hop Communication for Multi-Agent Reinforcement Learning\u003C\u002Fb>\u003C\u002Fi>, Wang et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAAMAS-2023.05-blue\" alt=\"AAMAS Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Task-Agnostic Contrastive Pre-Training for Inter-Agent Communication\u003C\u002Fb>\u003C\u002Fi>, Sun et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.02174\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAAMAS-2025.05-blue\" alt=\"AAMAS Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>AC2C: Adaptively Controlled Two-Hop Communication for Multi-Agent Reinforcement Learning\u003C\u002Fb>\u003C\u002Fi>, Xuefeng Wang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2302.12515\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.12-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>CAMEL: Communicative Agents for \"Mind\" Exploration of Large Language Model Society\u003C\u002Fb>\u003C\u002Fi>, Guohao Li et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2303.17760\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.11-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fcamel-ai\u002Fcamel\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fcamel-ai\u002Fcamel.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Context-aware Communication for Multi-agent Reinforcement Learning (CACOM)\u003C\u002Fb>\u003C\u002Fi>, Xinran Li et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.15600\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.07-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FLXXXXR\u002FCACOM\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FLXXXXR\u002FCACOM.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Task-Agnostic Contrastive Pre-Training for Inter-Agent Communication\u003C\u002Fb>\u003C\u002Fi>, Peihong Yu et al.\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Expressive Multi-Agent Communication via Identity-Aware Learning\u003C\u002Fb>\u003C\u002Fi>, Wei Du et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.07872\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.03-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>MAGIS: LLM-Based Multi-Agent Framework for GitHub Issue Resolution\u003C\u002Fb>\u003C\u002Fi>, Wei Tao et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2403.17927\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.06-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>AutoAgents: A Framework for Automatic Agent Generation\u003C\u002Fb>\u003C\u002Fi>, Guangyao Chen et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2309.17288\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FLink-AGI\u002FAutoAgents\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FLink-AGI\u002FAutoAgents.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>MDTeamGPT: A Self-Evolving LLM-based Multi-Agent Framework for Multi-Disciplinary Team Medical Consultation\u003C\u002Fb>\u003C\u002Fi>, Kai Chen et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2503.13856\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.03-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fkaichennj.github.io\u002FMDTeamGPT\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fkaichennj\u002FMDTeamGPT.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation\u003C\u002Fb>\u003C\u002Fi>, Wu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2308.08155\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.08-red\" alt=\"arXiv Badge\">\u003C\u002Fa> \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002Fautogen\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmicrosoft\u002Fautogen.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n\u003Cb>LLM-Enhanced Agent Communication\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>ProAgent: Building Proactive Cooperative AI with Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Ceyao Zhang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2308.11339\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.01-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fpku-proagent.github.io\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fpku-proagent\u002Fproagent.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Improving Factuality and Reasoning in Language Models through Multiagent Debate\u003C\u002Fb>\u003C\u002Fi>, Yilun Du et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2305.14325\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fcomposable-models.github.io\u002Fllm_debate\u002F\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fcomposable-models\u002Fllm_debate.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>ChatDev: Communicative Agents for Software Development\u003C\u002Fb>\u003C\u002Fi>, Chen Qian et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2307.07924\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.06-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FOpenBMB\u002FChatDev\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FOpenBMB\u002FChatDev.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Multi-Agent Incentive Communication via Decentralized Teammate Modeling\u003C\u002Fb>\u003C\u002Fi>, Nian Li et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.10436\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Ftsinghua-fib-lab\u002FACL24-EconAgent\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Ftsinghua-fib-lab\u002FACL24-EconAgent.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>AgentCoord: Visually Exploring Coordination Strategy for LLM-based Multi-Agent Collaboration\u003C\u002Fb>\u003C\u002Fi>, Bo Pan et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.11943\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FAgentCoord\u002FAgentCoord\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FAgentCoord\u002FAgentCoord.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Simulating Opinion Dynamics with Networks of LLM-based Agents\u003C\u002Fb>\u003C\u002Fi>, Yun-Shiuan Chuang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2311.09618\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fyunshiuan\u002Fllm-agent-opinion-dynamics\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fyunshiuan\u002Fllm-agent-opinion-dynamics.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework\u003C\u002Fb>\u003C\u002Fi>, Sirui Hong et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2308.00352\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.11-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fgeekan\u002FMetaGPT\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fgeekan\u002FMetaGPT.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Chain of Agents: Large Language Models Collaborating on Long-Context Tasks\u003C\u002Fb>\u003C\u002Fi>, Yusen Zhang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2406.02818\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.06-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Multi-Agent Incentive Communication via Decentralized Teammate Modeling\u003C\u002Fb>\u003C\u002Fi>, Lei Yuan et al.\u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.1609\u002Faaai.v36i9.21179\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAAAI-2022.06-blue\" alt=\"DOI Badge\">\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>ProAgent: Building Proactive Cooperative Agents with Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Zhang et al., \u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.1609\u002Faaai.v38i16.29710\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAAAI-2024.03-blue\" alt=\"AAAI Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FPKU-Alignment\u002FProAgent\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FPKU-Alignment\u002FProAgent.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Model Context Protocol (MCP)\u003C\u002Fb>\u003C\u002Fi>, Anthropic, \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmodelcontextprotocol\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FGitHub-2024-white\" alt=\"GitHub Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>CoMAS: Co-Evolving Multi-Agent Systems via Interaction Rewards\u003C\u002Fb>\u003C\u002Fi>, Xue et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2510.08529\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.10-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Achilles Heel of Distributed Multi-Agent Systems\u003C\u002Fb>\u003C\u002Fi>, Zhang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2504.07461\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n### 4. Tool Use and Function Calling\n\n#### Hosted Agent Tools and Computer Use\n\nThe frontier of tool use has shifted from static function schemas to **hosted tool runtimes**, **remote servers**, and **computer use interfaces**. In the agent era, tools are increasingly connected through platform-managed execution, approval flows, and UI-aware control loops rather than single-shot JSON calls.\n\n\u003Cb>Official Tooling and Computer Use References\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>OpenAI Tools Guide\u003C\u002Fb>\u003C\u002Fi>, OpenAI, \u003Ca href=\"https:\u002F\u002Fdevelopers.openai.com\u002Fapi\u002Fdocs\u002Fguides\u002Ftools\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpenAI-2026-blue\" alt=\"OpenAI Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Introducing Codex\u003C\u002Fb>\u003C\u002Fi>, OpenAI, \u003Ca href=\"https:\u002F\u002Fopenai.com\u002Findex\u002Fintroducing-codex\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpenAI-2025.05-blue\" alt=\"OpenAI Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Computer Use for Claude 3.5\u003C\u002Fb>\u003C\u002Fi>, Anthropic, \u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fnews\u002F3-5-models-and-computer-use\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAnthropic-2024.10-blue\" alt=\"Anthropic Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Google Vertex AI Agent Engine\u003C\u002Fb>\u003C\u002Fi>, Google, \u003Ca href=\"https:\u002F\u002Fcloud.google.com\u002Fvertex-ai\u002Fgenerative-ai\u002Fdocs\u002Fagent-engine\u002Foverview\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FGoogle-2026-blue\" alt=\"Google Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>OSWorld\u003C\u002Fb>\u003C\u002Fi>, Xie et al., \u003Ca href=\"https:\u002F\u002Fos-world.github.io\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FBenchmark-2026-orange\" alt=\"Benchmark Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Lumen\u003C\u002Fb>\u003C\u002Fi> — Vision-first browser agent with self-healing deterministic replay over CDP. Screenshot → model → action loop with multi-provider support (Anthropic, Google). \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fomxyz\u002Flumen\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fomxyz\u002Flumen.svg?style=social\" alt=\"GitHub stars\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Foundational Tool Learning\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Toolformer: Language Models Can Teach Themselves to Use Tools\u003C\u002Fb>\u003C\u002Fi>, Schick et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2302.04761\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNeurIPS-2023.09-blue\" alt=\"NeurIPS Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fxrsrke\u002Ftoolformer\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fxrsrke\u002Ftoolformer.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>ReAct: Synergizing Reasoning and Acting in Language Models\u003C\u002Fb>\u003C\u002Fi>, Yao et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.03629\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.10-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fysymyth\u002FReAct\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fysymyth\u002FReAct.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Augmented Language Models: a Survey\u003C\u002Fb>\u003C\u002Fi>, Qin et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2302.07842\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Tool Learning with Large Language Models: A Survey\u003C\u002Fb>\u003C\u002Fi>, Qu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.17935\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fquchangle1\u002FLLM-Tool-Survey\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fquchangle1\u002FLLM-Tool-Survey.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\u003Cb>Advanced Function Calling Systems\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Granite-Function Calling Model: Introducing Function Calling Abilities via Multi-task Learning of Granular Tasks\u003C\u002Fb>\u003C\u002Fi>, Smith et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.00121\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.06-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>HuggingGPT: Solving AI Tasks with ChatGPT and its Friends in Hugging Face\u003C\u002Fb>\u003C\u002Fi>, Shen et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2303.17580\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNeurIPS-2023.09-blue\" alt=\"NeurIPS Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fquchangle1\u002Fmicrosoft\u002FJARVIS\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmicrosoft\u002FJARVIS.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Enhancing Function-Calling Capabilities in LLMs: Strategies for Prompt Formats, Data Integration, and Multilingual Translation\u003C\u002Fb>\u003C\u002Fi>, Chen et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.01130\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNAACL-2025.04-blue\" alt=\"NAACL Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>A Real-World WebAgent for Complex Web Tasks\u003C\u002Fb>\u003C\u002Fi>, Zhai et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fhtml\u002F2307.12856v4\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.03-red\" alt=\"arXiv Badge\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\u003Cb>Multi-Agent Function Calling\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>ToolACE: Winning the Points of LLM Function Calling\u003C\u002Fb>\u003C\u002Fi>, Zhang et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpenReview-2025.03-orange\" alt=\"OpenReview Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Berkeley Function Leaderboard (BFCL): Evaluating Function-Calling Abilities\u003C\u002Fb>\u003C\u002Fi>, Various, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FBenchmark-2024-orange\" alt=\"Benchmark Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FShishirPatil\u002Fgorilla\u002Ftree\u002Fmain\u002Fberkeley-function-call-leaderboard\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FShishirPatil\u002Fgorilla.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n---\n\n## 📊 Evaluation Paradigms for Context-Driven Systems\n\n### Context Quality Assessment\n\n\u003Cb>Foundational Long-Context Benchmarks\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>RULER: What's the Real Context Size of Your Long-Context Language Models?\u003C\u002Fb>\u003C\u002Fi>, Cheng-Ping Hsieh et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.06654\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FCOLM-2024.07-blue\" alt=\"COLM Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FRULER\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FNVIDIA\u002FRULER.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding\u003C\u002Fb>\u003C\u002Fi>, Bai et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FACL-2024.08-blue\" alt=\"ACL Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>∞BENCH: Extending Long Context Evaluation Beyond 100K Tokens\u003C\u002Fb>\u003C\u002Fi>, Zhang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.13718\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FACL-2024.08-blue\" alt=\"ACL Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FTHUDM\u002FLongBench\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FTHUDM\u002FLongBench.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>VL-ICL Bench: The Devil in the Details of Multimodal In-Context Learning\u003C\u002Fb>\u003C\u002Fi>, Zong et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2025.01-blue\" alt=\"ICLR Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fys-zong\u002FVL-ICL\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fys-zong\u002FVL-ICL.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\u003Cb>Multimodal and Specialized Evaluation\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>MultiModal Needle in a Haystack: Benchmarking Long-Context Capability of Multimodal Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Wang et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNAACL-2025.04-blue\" alt=\"NAACL Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FWang-ML-Lab\u002Fmultimodal-needle-in-a-haystack\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FWang-ML-Lab\u002Fmultimodal-needle-in-a-haystack.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Contextualized Topic Coherence (CTC) Metrics\u003C\u002Fb>\u003C\u002Fi>, Rahimi et al., \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FACL-2024.03-blue\" alt=\"ACL Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FhamedR96\u002FCTC\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FhamedR96\u002FCTC.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>BBScore: A Brownian Bridge Based Metric for Assessing Text Coherence\u003C\u002Fb>\u003C\u002Fi>, Sheng et al., \u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.1609\u002Faaai.v38i13.29414\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAAAI-2024.03-blue\" alt=\"AAAI Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fzcsheng95\u002FBBScore\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fzcsheng95\u002FBBScore.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\u003Cb>RAG and Generation Evaluation\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Evaluation of Retrieval-Augmented Generation: A Survey\u003C\u002Fb>\u003C\u002Fi>, Li et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.07437\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Ragas: Automated Evaluation of Retrieval Augmented Generation\u003C\u002Fb>\u003C\u002Fi>, Espinosa-Anke et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2309.15217\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.09-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Human Evaluation Protocol for Generative AI Chatbots in Clinical Microbiology\u003C\u002Fb>\u003C\u002Fi>, Griego-Herrera et al., \u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.1371\u002Fjournal.pone.0300487\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FPLOS-2024.03-blue\" alt=\"PLOS Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n### Benchmarking Context Engineering\n\n\u003Cb>Synthetic vs. Realistic Evaluation\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Needle-in-a-Haystack (NIAH) and Synthetic Benchmarks\u003C\u002Fb>\u003C\u002Fi>, Research Area 2023-2024, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FBenchmark-2024-orange\" alt=\"Benchmark Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fgkamradt\u002FLLMTest_NeedleInAHaystack\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fgkamradt\u002FLLMTest_NeedleInAHaystack.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>ZeroSCROLLS: Realistic Natural Language Tasks\u003C\u002Fb>\u003C\u002Fi>, Benchmark 2023-2024, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FBenchmark-2024-orange\" alt=\"Benchmark Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Ftau-nlp\u002Fzero_scrolls\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Ftau-nlp\u002Fzero_scrolls.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>InfiniteBench: 100K+ Token Evaluation\u003C\u002Fb>\u003C\u002Fi>, Benchmark 2024, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FBenchmark-2024-orange\" alt=\"Benchmark Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FOpenBMB\u002FInfiniteBench\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FOpenBMB\u002FInfiniteBench.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Agent-Pro: Learning to Evolve Coder Agents via Proposal-based Programming\u003C\u002Fb>\u003C\u002Fi>, Zhang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.17574\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.05-red\" alt=\"arXiv Badge\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>GenoTEX: An LLM Agent Benchmark for Automated Gene Expression Data Analysis\u003C\u002Fb>\u003C\u002Fi>, Liu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2406.15341\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FMLCB-2025.06-blue\" alt=\"MLCB Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FLiu-Hy\u002FGenoTEX\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FLiu-Hy\u002FGenoTEX.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n### Agent Observability and Telemetry\n\nLong-running agent systems need more than offline benchmark scores. They require trace-level visibility into plans, tool calls, memory reads and writes, approvals, retries, and failure modes. Observability is increasingly the verification layer for context engineering in production.\n\n\u003Cb>Observability and Telemetry References\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>LangSmith Observability Quickstart\u003C\u002Fb>\u003C\u002Fi>, LangChain, \u003Ca href=\"https:\u002F\u002Fdocs.langchain.com\u002Flangsmith\u002Fobservability-quickstart\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLangChain-2026-blue\" alt=\"LangChain Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>OpenTelemetry Semantic Conventions for Generative AI\u003C\u002Fb>\u003C\u002Fi>, OpenTelemetry, \u003Ca href=\"https:\u002F\u002Fopentelemetry.io\u002Fdocs\u002Fspecs\u002Fsemconv\u002Fgen-ai\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpenTelemetry-2026-blue\" alt=\"OpenTelemetry Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Google ADK Evaluation and Observability\u003C\u002Fb>\u003C\u002Fi>, Google, \u003Ca href=\"https:\u002F\u002Fgoogle.github.io\u002Fadk-docs\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FGoogle-2026-blue\" alt=\"Google Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>OpenAI Agents and Tools\u003C\u002Fb>\u003C\u002Fi>, OpenAI, \u003Ca href=\"https:\u002F\u002Fplatform.openai.com\u002Fdocs\u002Fguides\u002Fagents\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpenAI-2026-blue\" alt=\"OpenAI Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n---\n\n## 🚀 Applications and Systems\n\n### Complex Research Systems\n\n\u003Cb>Hypothesis Generation and Data-Driven Discovery\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Hypothesis Generation with Large Language Models\u003C\u002Fb>\u003C\u002Fi>, Liu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.04326\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FChicagoHAI\u002Fhypothesis-generation\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FChicagoHAI\u002Fhypothesis-generation.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>GFlowNets for AI-Driven Scientific Discovery\u003C\u002Fb>\u003C\u002Fi>, Jain et al., \u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.1039\u002FD3DD00002H\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FDigital_Discovery-2023.06-blue\" alt=\"Digital Discovery Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Literature Meets Data: A Synergistic Approach to Hypothesis Generation\u003C\u002Fb>\u003C\u002Fi>, Liu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.17309\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.10-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FChicagoHAI\u002Fhypothesis-generation\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FChicagoHAI\u002Fhypothesis-generation.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Machine Learning for Hypothesis Generation in Biology and Medicine\u003C\u002Fb>\u003C\u002Fi>, FieldSHIFT Team, \u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.1039\u002FD3DD00185G\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FDigital_Discovery-2024.02-blue\" alt=\"Digital Discovery Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Automated Scientific Discovery\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery\u003C\u002Fb>\u003C\u002Fi>, Lu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2408.06292\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.08-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FSakanaAI\u002FAI-Scientist\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FSakanaAI\u002FAI-Scientist.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Automating Psychological Hypothesis Generation with AI\u003C\u002Fb>\u003C\u002Fi>, Johnson et al., \u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.1038\u002Fs41599-024-03407-5\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNature-2024.07-blue\" alt=\"Nature Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Can Large Language Models Replace Humans in Systematic Reviews?\u003C\u002Fb>\u003C\u002Fi>, Khraisha et al., \u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.1002\u002Fjrsm.1715\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FResearch_Synthesis-2024-blue\" alt=\"Research Synthesis Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Solving Olympiad Geometry without Human Demonstrations\u003C\u002Fb>\u003C\u002Fi>, Trinh et al., \u003Ca href=\"https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41586-023-06747-5\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNature-2024.01-blue\" alt=\"Nature Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>GenoMAS: A Multi-Agent Framework for Scientific Discovery via Code-Driven Gene Expression Analysis\u003C\u002Fb>\u003C\u002Fi>, Liu et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.21035\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.07-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FLiu-Hy\u002FGenoMAS\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FLiu-Hy\u002FGenoMAS.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>aiXiv: A Next-Generation Open Access Ecosystem for Scientific Discovery Generated by AI Scientists\u003C\u002Fb>\u003C\u002Fi>, Zhang et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2508.15126\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.08-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Faixiv-org\u002FaiXiv\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Faixiv-org\u002FaiXiv.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>AI for Science Integration and Future Directions\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>AI for Science 2025: Convergence of AI Innovation and Scientific Discovery\u003C\u002Fb>\u003C\u002Fi>, Fink et al., \u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.1038\u002Fd41573-025-00161-3\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNature-2025.05-blue\" alt=\"Nature Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Towards Scientific Discovery with Generative AI: Progress, Opportunities, and Challenges\u003C\u002Fb>\u003C\u002Fi>, Anonymous et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.11427\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.12-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Deep Research Applications\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Accelerating scientific discovery with AI\u003C\u002Fb>\u003C\u002Fi>, MIT News, \u003Ca href=\"https:\u002F\u002Fnews.mit.edu\u002F2025\u002Ffuturehouse-accelerates-scientific-discovery-with-ai-0630\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FMIT-2025.06-blue\" alt=\"MIT Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Accelerating scientific breakthroughs with an AI co-scientist\u003C\u002Fb>\u003C\u002Fi>, Google Research, \u003Ca href=\"https:\u002F\u002Fresearch.google\u002Fblog\u002Faccelerating-scientific-breakthroughs-with-an-ai-co-scientist\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FGoogle-2024-blue\" alt=\"Google Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Bridging AI and Science: Implications from a Large-Scale Literature Analysis of AI4Science\u003C\u002Fb>\u003C\u002Fi>, Various, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.09628\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.12-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fcharles-pyj\u002FBridging-AI-and-Science\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fcharles-pyj\u002FBridging-AI-and-Science.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>AI for scientific discovery\u003C\u002Fb>\u003C\u002Fi>, World Economic Forum, \u003Ca href=\"https:\u002F\u002Fwww.weforum.org\u002Fpublications\u002Ftop-10-emerging-technologies-2024\u002Fin-full\u002F1-ai-for-scientific-discovery\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FWEF-2024-blue\" alt=\"WEF Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\n### Production Systems\n\n\u003Cb>Context Engineering as a Core Discipline\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>From Prompt Craft to System Design: Context Engineering as a Core Discipline for AI-Driven Delivery\u003C\u002Fb>\u003C\u002Fi>, Forte Group Team, \u003Ca href=\"https:\u002F\u002Ffortegrp.com\u002Finsights\u002Fcontext-engineering-as-a-core-discipline-for-ai-driven-delivery\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FForte-2025.07-blue\" alt=\"Forte Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Context Engineering: A Framework for Enterprise AI Operations\u003C\u002Fb>\u003C\u002Fi>, Shelly Palmer, \u003Ca href=\"https:\u002F\u002Fshellypalmer.com\u002F2025\u002F06\u002Fcontext-engineering-a-framework-for-enterprise-ai-operations\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FShellyPalmer-2025.06-blue\" alt=\"ShellyPalmer Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>How MCP Handles Context Management in High-Throughput Scenarios\u003C\u002Fb>\u003C\u002Fi>, Portkey.ai Team, \u003Ca href=\"https:\u002F\u002Fportkey.ai\u002Fblog\u002Fmodel-context-protocol-context-management-in-high-throughput\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FPortkey-2025.03-blue\" alt=\"Portkey Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Enterprise AI Case Studies\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Case Study: JPMorgan's COiN Platform – Agentic AI for Financial Analysis\u003C\u002Fb>\u003C\u002Fi>, AI Mindset Research, \u003Ca href=\"https:\u002F\u002Fwww.ai-mindset.ai\u002Fenterprise-ai-case-studies#JPMorgan\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FBanking-2025.02-green\" alt=\"Banking Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Case Study: EY's Agentic AI Integration in Microsoft 365 Copilot\u003C\u002Fb>\u003C\u002Fi>, AI Mindset Research, \u003Ca href=\"https:\u002F\u002Fwww.ai-mindset.ai\u002Fenterprise-ai-case-studies#EY\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FProfessional_Services-2025.02-green\" alt=\"Professional Services Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Context Is Everything: The Massive Shift Making AI Actually Work in the Real World\u003C\u002Fb>\u003C\u002Fi>, Phil Mora, \u003Ca href=\"https:\u002F\u002Fwww.philmora.com\u002Fthe-big-picture\u002Fcontext-is-everything-the-massive-shift-making-ai-actually-work-in-the-real-world\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FCross_Industry-2025.06-green\" alt=\"Cross Industry Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Enterprise Applications and Infrastructure\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>The Context Layer for Enterprise RAG Applications\u003C\u002Fb>\u003C\u002Fi>, Contextual AI Team, \u003Ca href=\"https:\u002F\u002Fcontextual.ai\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FContextual_AI-2025.07-blue\" alt=\"Contextual AI Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Navigating AI Model Deployment: Challenges and Solutions\u003C\u002Fb>\u003C\u002Fi>, Dean Lancaster, \u003Ca href=\"https:\u002F\u002Fwww.linkedin.com\u002Fpulse\u002Ffrom-poc-production-overcoming-ai-deployment-ensuring-dean-lancaster-fmtoe\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLinkedIn-2025.03-blue\" alt=\"LinkedIn Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>2024: The State of Generative AI in the Enterprise\u003C\u002Fb>\u003C\u002Fi>, Menlo Ventures, \u003Ca href=\"https:\u002F\u002Fmenlovc.com\u002F2024-the-state-of-generative-ai-in-the-enterprise\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FReport-2024-blue\" alt=\"Report Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>How 100 Enterprise CIOs Are Building and Buying Gen AI in 2025\u003C\u002Fb>\u003C\u002Fi>, Andreessen Horowitz, \u003Ca href=\"https:\u002F\u002Fa16z.com\u002Fai-enterprise-2025\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fa16z-2025-blue\" alt=\"a16z Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>Developer Tools with Context Engineering\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Autohand Code CLI: Autonomous Coding Agent with Semantic Search, Memory, and Context Management\u003C\u002Fb>\u003C\u002Fi>, Autohand AI, \u003Ca href=\"https:\u002F\u002Fwww.autohand.ai\u002Fcode\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FTool-2025-green\" alt=\"Tool Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fautohandai\u002Fcode-cli\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fautohandai\u002Fcode-cli.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n#### Coding Agents and Project Memory\n\nCoding agents are one of the clearest production settings in which context engineering becomes agent engineering. Here, context is no longer just a prompt: it becomes repository instructions, project memory, task plans, file diffs, test results, and tool traces.\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Introducing Codex\u003C\u002Fb>\u003C\u002Fi>, OpenAI, \u003Ca href=\"https:\u002F\u002Fopenai.com\u002Findex\u002Fintroducing-codex\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpenAI-2025.05-blue\" alt=\"OpenAI Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Claude Code Memory\u003C\u002Fb>\u003C\u002Fi>, Anthropic, \u003Ca href=\"https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fdocs\u002Fclaude-code\u002Fmemory\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAnthropic-2026-blue\" alt=\"Anthropic Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Claude Code Subagents\u003C\u002Fb>\u003C\u002Fi>, Anthropic, \u003Ca href=\"https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fdocs\u002Fclaude-code\u002Fsub-agents\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAnthropic-2026-blue\" alt=\"Anthropic Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Letta Memory Blocks\u003C\u002Fb>\u003C\u002Fi>, Letta, \u003Ca href=\"https:\u002F\u002Fdocs.letta.com\u002Fguides\u002Fcore-concepts\u002Fmemory\u002Fmemory-blocks\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLetta-2026-blue\" alt=\"Letta Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LangChain Deep Agents\u003C\u002Fb>\u003C\u002Fi>, LangChain, \u003Ca href=\"https:\u002F\u002Fdocs.langchain.com\u002Foss\u002Fpython\u002Fdeepagents\u002Foverview\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLangChain-2026-blue\" alt=\"LangChain Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n#### Platform Stacks and Hosted Agent Runtimes\n\nThe production ecosystem is increasingly organized around full agent stacks rather than isolated models or prompts. These stacks combine tools, memory, runtime orchestration, sessions, observability, and interoperability in a single platform surface.\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>OpenAI Agents Guide\u003C\u002Fb>\u003C\u002Fi>, OpenAI, \u003Ca href=\"https:\u002F\u002Fplatform.openai.com\u002Fdocs\u002Fguides\u002Fagents\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpenAI-2026-blue\" alt=\"OpenAI Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Google Agent Development Kit (ADK)\u003C\u002Fb>\u003C\u002Fi>, Google, \u003Ca href=\"https:\u002F\u002Fgoogle.github.io\u002Fadk-docs\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FGoogle-2026-blue\" alt=\"Google Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Vertex AI Agent Engine\u003C\u002Fb>\u003C\u002Fi>, Google, \u003Ca href=\"https:\u002F\u002Fcloud.google.com\u002Fvertex-ai\u002Fgenerative-ai\u002Fdocs\u002Fagent-engine\u002Foverview\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FGoogle-2026-blue\" alt=\"Google Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LangGraph Memory Overview\u003C\u002Fb>\u003C\u002Fi>, LangChain, \u003Ca href=\"https:\u002F\u002Fdocs.langchain.com\u002Foss\u002Fjavascript\u002Flanggraph\u002Fmemory\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLangChain-2026-blue\" alt=\"LangChain Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Microsoft Agent Framework\u003C\u002Fb>\u003C\u002Fi>, Microsoft, \u003Ca href=\"https:\u002F\u002Flearn.microsoft.com\u002Fen-us\u002Fagent-framework\u002Fuser-guide\u002Foverview\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FMicrosoft-2026-blue\" alt=\"Microsoft Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n---\n\n## 🔮 Limitations and Future Directions\n\n### Current Limitations\n\n1. **Context Window Constraints**: Despite improvements, context length remains a bottleneck\n2. **Computational Overhead**: Processing large contexts requires significant resources\n3. **Context Coherence**: Maintaining coherence across extended contexts\n4. **Dynamic Adaptation**: Real-time context updating challenges\n\n### Future Research Directions\n\n1. **Infinite Context**: Developing truly unlimited context capabilities\n2. **Context Compression**: Efficient representation of large contexts\n3. **Multimodal Integration**: Seamless integration of diverse data types\n4. **Adaptive Context**: Self-optimizing context management\n5. **Context Privacy**: Securing sensitive information in context pipelines\n\n---\n\n## 🤝 Contributing\n\nWe welcome contributions to this survey! Please follow these guidelines:\n\n1. **Fork** the repository\n2. **Create** a feature branch\n3. **Add** relevant papers with proper formatting\n4. **Submit** a pull request with a clear description\n\n### Paper Formatting Guidelines\n\n```markdown\n\u003Cli>\u003Ci>\u003Cb>Paper Title\u003C\u002Fb>\u003C\u002Fi>, Author et al., \u003Ca href=\"URL\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FSOURCE-YEAR.MM-COLOR\" alt=\"SOURCE Badge\">\u003C\u002Fa>\u003C\u002Fli>\n```\n\n### Badge Colors\n- ![arXiv Badge](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-red) `red` for arXiv papers\n- ![PDF Badge](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FPDF-blue) `blue` for conference\u002Fjournal papers\n- ![GitHub Badge](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FGitHub-white) `white` for GitHub repositories\n- ![HuggingFace Badge](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FHuggingFace-yellow) `yellow` for HuggingFace resources\n\n---\n\n## 📄 License\n\nThis project is licensed under the MIT License - see the [LICENSE](LICENSE) file for details.\n\n---\n\n## 📑 Citation\n\nIf you find this survey helpful in your research, please consider citing:\n\n```bibtex\n@misc{mei2025surveycontextengineeringlarge,\n      title={A Survey of Context Engineering for Large Language Models}, \n      author={Lingrui Mei and Jiayu Yao and Yuyao Ge and Yiwei Wang and Baolong Bi and Yujun Cai and Jiazhi Liu and Mingyu Li and Zhong-Zhi Li and Duzhen Zhang and Chenlin Zhou and Jiayi Mao and Tianze Xia and Jiafeng Guo and Shenghua Liu},\n      year={2025},\n      eprint={2507.13334},\n      archivePrefix={arXiv},\n      primaryClass={cs.CL},\n      url={https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.13334}, \n}\n```\n\n---\n\n## ⚠️ Disclaimer\n\nThis project is **ongoing** and continuously evolving. While we strive for accuracy and completeness, there may be errors, omissions, or outdated information. We welcome corrections, suggestions, and contributions from the community. Please stay tuned for regular updates and improvements.\n\n## 📧 Contact\n\nFor questions, suggestions, or collaboration opportunities, please feel free to reach out:\n\n**Lingrui Mei**  \n📧 Email: [meilingrui22@mails.ucas.ac.cn](mailto:meilingrui22@mails.ucas.ac.cn)\n\nYou can also open an issue in this repository for general discussions and suggestions.\n\n---\n\n## 🙏 Acknowledgments\n\nThis survey builds upon the foundational work of the AI research community. We thank all researchers contributing to the advancement of context engineering and large language models.\n\n---\n\n## Star History\n\n**Star ⭐ this repository if you find it helpful!**\n\n[![Star History Chart](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FMeirtz_Awesome-Context-Engineering_readme_bc8c8bff666e.png)](https:\u002F\u002Fwww.star-history.com\u002F#Meirtz\u002FAwesome-Context-Engineering&Date)\n\n---\n\n## 📖 Our Paper\n\n**A Survey of Context Engineering for Large Language Models**\n\n- **arXiv**: https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.13334\n- **Hugging Face Papers**: https:\u002F\u002Fhuggingface.co\u002Fpapers\u002F2507.13334\n\nThis comprehensive survey provides the latest academic insights and theoretical foundations for context engineering in large language models.\n","# 令人惊叹的上下文工程\n\n\u003Cdiv align=\"center\">\n  \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FMeirtz_Awesome-Context-Engineering_readme_f88855f12e97.png\" alt=\"Awesome Context Engineering Cover\" width=\"800\"\u002F>\n\u003C\u002Fdiv>\n\n## 💬 加入我们的社区\n\n\u003Cdiv align=\"center\">\n  \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FMeirtz_Awesome-Context-Engineering_readme_378aa8ccb7ca.png\" alt=\"WeChat Group\" width=\"200\"\u002F>\n  \u003Cp>\u003Cstrong>加入我们的微信群，参与讨论和获取最新动态！\u003C\u002Fstrong>\u003C\u002Fp>\n  \u003Cp>\u003Ca href=\"https:\u002F\u002Fdiscord.gg\u002Ffsqs3Ybh\">\u003Cstrong>加入我们的 Discord 服务器\u003C\u002Fstrong>\u003C\u002Fa>\u003C\u002Fp>\n\u003C\u002Fdiv>\n\n[![Awesome](https:\u002F\u002Fawesome.re\u002Fbadge.svg)](https:\u002F\u002Fawesome.re)\n[![License: MIT](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLicense-MIT-yellow.svg)](https:\u002F\u002Fopensource.org\u002Flicenses\u002FMIT)\n[![PRs Welcome](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FPRs-welcome-brightgreen.svg?style=flat-square)](http:\u002F\u002Fmakeapullrequest.com)\n[![Paper](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FPaper-Published-green.svg)](https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.13334)\n\n> 📄 **我们关于上下文工程的综合综述论文现已发表！** 欢迎查阅我们的最新学术见解和理论基础。\n\n这是一份关于 **上下文工程** 的全面综述与资源汇编——从静态提示逐步演进到动态、上下文感知的人工智能系统，并进一步发展为 **代理运行时、记忆系统、协议、编码代理以及可观测性栈**。\n\n## 📧 联系方式\n\n如有任何问题、建议或合作机会，请随时联系我们：\n\n**梅凌睿**  \n📧 邮箱：  [meilingrui25b@ict.ac.cn](mailto:meilingrui25b@ict.ac.cn) 或 [meilingrui22@mails.ucas.ac.cn](mailto:meilingrui22@mails.ucas.ac.cn)\n\n**我在论文初稿中写错了邮箱地址！！** 您也可以在此仓库中提交议题，进行一般性讨论和建议。\n\n---\n\n## 📰 最新消息\n\n- **[2025.07.17]** 🔥🔥 我们的论文现已发表！请查看 [\"大型语言模型的上下文工程综述\"](https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.13334) ，发表于 [arXiv](https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.13334) 和 [Hugging Face Papers](https:\u002F\u002Fhuggingface.co\u002Fpapers\u002F2507.13334)\n- **[2025.07.03]** 仓库初始化，包含全面的大纲\n- **[2025.07.03]** 根据现代上下文工程范式建立了综述框架\n\n---\n\n## 🎯 引言\n\n在大型语言模型（LLMs）时代，静态提示的局限性日益凸显。**上下文工程** 是应对 LLM 不确定性并实现生产级 AI 部署的自然演进方向。与传统的提示工程不同，上下文工程涵盖了推理时提供给 LLM 的完整信息载荷，包括完成合理任务所需的所有结构化信息组件。\n\n本仓库旨在作为上下文工程技术、方法论及应用的全面综述。\n\n---\n\n## 🧭 2026 年代理时代更新\n\n### 从上下文工程到代理工程\n\n截至 **2026 年 3 月**，上下文工程仍然是一个有用且必要的概念，但它已不再是全部。重心已经从“如何打包最佳提示”转移到 **代理系统如何管理运行时状态、记忆、工具、协议、审批流程以及长周期执行**。实际上，如今上下文工程已融入更广泛的体系中，其中包括 **代理框架**、**互操作性协议**、**用于编码代理的项目记忆**以及 **以追踪为核心的可观测性**。\n\n### 本仓库目前涵盖的内容\n\n本仓库仍保留了原有的关于长上下文、RAG、记忆、代理间通信、工具使用、评估及应用的综述结构。与此同时，本 README 正在重新组织，以更好地反映 **代理时代**，新增以下内容：\n\n- **代理框架和运行时系统**，用于规划、子代理、检查点、沙盒环境以及人工审批流程\n- **生产环境中的上下文管理**，通过压缩、缓存、基于工件的上下文以及按范围加载指令等方式实现\n- **记忆工件与可移植性**，包括持久化记忆、记忆交换格式、角色封装以及项目记忆\n- **开放协议**，如 MCP、A2A、AG-UI、ACP 以及可移植的代理模式\n- **编码代理与计算机使用**，作为当前上下文工程最显著的生产应用场景\n- **评估、可观测性与遥测**，针对长期运行的代理系统，而不仅仅是静态基准测试\n\n### 2026 年主题阅读指南\n\n主要关注 2026 年转变的读者，可直接跳转至以下扩展章节：\n\n- **代理框架和运行时系统**，灵感来源于 [Anthropic 的高效代理指南](https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Fbuilding-effective-agents)、[OpenAI 的代理与工具文档](https:\u002F\u002Fplatform.openai.com\u002Fdocs\u002Fguides\u002Fagents)、[Google ADK](https:\u002F\u002Fgoogle.github.io\u002Fadk-docs\u002F) 以及 [LangChain Deep Agents](https:\u002F\u002Fdocs.langchain.com\u002Foss\u002Fpython\u002Fdeepagents\u002Foverview)\n- **开放协议与互操作性**，包括 [Model Context Protocol](https:\u002F\u002Fmodelcontextprotocol.io\u002Fspecification\u002F2025-06-18)、[A2A](https:\u002F\u002Fa2a-protocol.org\u002Flatest\u002F)、[AG-UI](https:\u002F\u002Fdocs.ag-ui.com\u002F) 以及 [AgentSchema](https:\u002F\u002Fmicrosoft.github.io\u002FAgentSchema\u002F)\n- **编码代理和项目记忆**，包括 [OpenAI Codex](https:\u002F\u002Fopenai.com\u002Findex\u002Fintroducing-codex\u002F)、[Claude Code 记忆](https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fdocs\u002Fclaude-code\u002Fmemory) 以及 [Letta 记忆块](https:\u002F\u002Fdocs.letta.com\u002Fguides\u002Fcore-concepts\u002Fmemory\u002Fmemory-blocks)\n- **评估和可观测性**，包括 [LangSmith 可观测性](https:\u002F\u002Fdocs.langchain.com\u002Flangsmith\u002Fobservability-quickstart) 和 [OpenTelemetry GenAI 语义规范](https:\u002F\u002Fopentelemetry.io\u002Fdocs\u002Fspecs\u002Fsemconv\u002Fgen-ai\u002F)\n\n---\n\n## 📚 目录\n\n- [Awesome Context Engineering](#awesome-context-engineering)\n  - [💬 加入我们的社区](#-join-our-community)\n  - [📧 联系我们](#-contact)\n  - [📰 新闻](#-news)\n  - [🎯 简介](#-introduction)\n  - [🧭 2026年智能体时代更新](#-2026-agent-era-update)\n    - [从上下文工程到智能体工程](#from-context-engineering-to-agent-engineering)\n    - [本仓库目前涵盖的内容](#what-this-repository-now-covers)\n    - [2026年主题阅读指南](#reading-guide-for-2026-topics)\n  - [📚 目录](#-table-of-contents)\n  - [🔗 相关综述](#-related-survey)\n  - [🏗️ 上下文工程的定义](#️-definition-of-context-engineering)\n    - [LLM生成](#llm-generation)\n    - [上下文的定义](#definition-of-context)\n    - [上下文工程的定义](#definition-of-context-engineering)\n    - [动态上下文编排](#dynamic-context-orchestration)\n    - [数学原理](#mathematical-principles)\n    - [理论框架：贝叶斯上下文推理](#theoretical-framework-bayesian-context-inference)\n    - [比较](#comparison)\n  - [🌐 相关博客](#-related-blogs)\n    - [社交媒体与演讲](#social-media--talks)\n  - [🤔 为什么是上下文工程？](#-why-context-engineering)\n    - [范式转变：从战术到战略](#the-paradigm-shift-from-tactical-to-strategic)\n    - [1. 当前方法的基本挑战](#1-fundamental-challenges-with-current-approaches)\n      - [人类意图沟通的挑战](#human-intent-communication-challenges)\n      - [复杂知识需求](#complex-knowledge-requirements)\n      - [可靠性和可信度问题](#reliability-and-trustworthiness-issues)\n    - [2. 静态提示的局限性](#2-limitations-of-static-prompting)\n      - [从字符串到系统](#from-strings-to-systems)\n      - [“电影制作”类比](#the-movie-production-analogy)\n    - [3. 企业与生产环境的需求](#3-enterprise-and-production-requirements)\n      - [上下文故障成为新的瓶颈](#context-failures-are-the-new-bottleneck)\n      - [超越简单任务的可扩展性](#scalability-beyond-simple-tasks)\n      - [可靠性和一致性](#reliability-and-consistency)\n      - [经济与运营效率](#economic-and-operational-efficiency)\n    - [4. 认知科学与信息科学的基础](#4-cognitive-and-information-science-foundations)\n      - [人工具身化](#artificial-embodiment)\n      - [大规模信息检索](#information-retrieval-at-scale)\n    - [5. AI系统架构的未来](#5-the-future-of-ai-system-architecture)\n  - [🔧 组件、技术和架构](#-components-techniques-and-architectures)\n    - [上下文扩展](#context-scaling)\n    - [生产环境中的上下文管理](#context-management-in-production)\n    - [结构化数据集成](#structured-data-integration)\n    - [自生成上下文](#self-generated-context)\n  - [🛠️ 实现与挑战](#️-implementation-and-challenges)\n    - [0. 智能体框架与运行时系统](#0-agent-harnesses-and-runtime-systems)\n    - [1. 检索增强生成（RAG）](#1-retrieval-augmented-generation-rag)\n    - [2. 内存系统](#2-memory-systems)\n      - [运行时内存设计模式](#runtime-memory-design-patterns)\n      - [项目记忆与指令工件](#project-memory-and-instruction-artifacts)\n    - [3. 智能体通信](#3-agent-communication)\n      - [开放的智能体协议与互操作性](#open-agent-protocols-and-interoperability)\n    - [4. 工具使用与函数调用](#4-tool-use-and-function-calling)\n      - [托管智能体工具与计算机使用](#hosted-agent-tools-and-computer-use)\n  - [📊 基于上下文驱动系统的评估范式](#-evaluation-paradigms-for-context-driven-systems)\n    - [上下文质量评估](#context-quality-assessment)\n    - [上下文工程的基准测试](#benchmarking-context-engineering)\n    - [智能体可观测性与遥测](#agent-observability-and-telemetry)\n  - [🚀 应用与系统](#-applications-and-systems)\n    - [复杂研究系统](#complex-research-systems)\n    - [生产系统](#production-systems)\n      - [编码智能体与项目记忆](#coding-agents-and-project-memory)\n      - [平台栈与托管智能体运行时](#platform-stacks-and-hosted-agent-runtimes)\n  - [🔮 局限性与未来方向](#-limitations-and-future-directions)\n    - [当前局限性](#current-limitations)\n    - [未来研究方向](#future-research-directions)\n  - [🤝 贡献](#-contributing)\n    - [论文格式指南](#paper-formatting-guidelines)\n    - [徽章颜色](#badge-colors)\n  - [📄 许可证](#-license)\n  - [📑 引用](#-citation)\n  - [⚠️ 免责声明](#️-disclaimer)\n  - [📧 联系我们](#-contact-1)\n  - [🙏 致谢](#-acknowledgments)\n  - [星标历史](#star-history)\n  - [📖 我们的论文](#-our-paper)\n\n---\n\n## 🔗 相关综述\n\n\u003Cb>通用AI综述论文\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>大型语言模型综述\u003C\u002Fb>\u003C\u002Fi>, Zhao等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2303.18223\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.03-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FRUCAIBox\u002FLLMSurvey\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FRUCAIBox\u002FLLMSurvey.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>提示报告：提示工程技巧的系统性综述\u003C\u002Fb>\u003C\u002Fi>, Schulhoff等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2406.06608\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.02-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Ftrigaten\u002FThe_Prompt_Report\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Ftrigaten\u002FThe_Prompt_Report.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>大型语言模型中提示工程的系统性综述：技术与应用\u003C\u002Fb>\u003C\u002Fi>, Sahoo等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.07927\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.03-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>视觉-语言基础模型上的提示工程系统性综述\u003C\u002Fb>\u003C\u002Fi>, Gao等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2307.12980\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.07-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FJindongGu\u002FAwesome-Prompting-on-Vision-Language-Model\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FJindongGu\u002FAwesome-Prompting-on-Vision-Language-Model.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>上下文与推理\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>上下文学习综述\u003C\u002Fb>\u003C\u002Fi>, Dong 等人，\u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.18653\u002Fv1\u002F2024.emnlp-main.64\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FEMNLP-2024.11-blue\" alt=\"EMNLP徽章\">\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fdqxiu\u002FICL_PaperList\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fdqxiu\u002FICL_PaperList.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>上下文学习之谜：解释与分析的全面综述\u003C\u002Fb>\u003C\u002Fi>, Zhou 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2311.00237\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.10-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fzyxnlp\u002FICL-Interpretation-Analysis-Resources\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fzyxnlp\u002FICL-Interpretation-Analysis-Resources.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>检索增强生成（RAG）的全面综述：演进、现状与未来方向\u003C\u002Fb>\u003C\u002Fi>, Gupta 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.12837\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.10-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>大型语言模型的检索增强生成：综述\u003C\u002Fb>\u003C\u002Fi>, Gao 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.10997\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.03-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FTongji-KGLLM\u002FRAG-Survey\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FTongji-KGLLM\u002FRAG-Survey.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>面向知识的检索增强生成综述\u003C\u002Fb>\u003C\u002Fi>, Cheng 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2503.10677\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.02-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FUSTCAGI\u002FAwesome-Papers-Retrieval-Augmented-Generation\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FUSTCAGI\u002FAwesome-Papers-Retrieval-Augmented-Generation.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>记忆系统与上下文持久性\u003C\u002Fb>\n\n\u003Cb>综述\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>基于大型语言模型的智能体记忆机制综述\u003C\u002Fb>\u003C\u002Fi>, Zhang 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.13501\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fnuster1128\u002FLLM_Agent_Memory_Survey\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fnuster1128\u002FLLM_Agent_Memory_Survey.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>记忆增强型神经网络综述：从认知洞察到人工智能应用\u003C\u002Fb>\u003C\u002Fi>, Khosla 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.06141\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.12-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>从人类记忆到AI记忆：大语言模型时代下的记忆机制综述\u003C\u002Fb>\u003C\u002Fi>, Wu 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2504.15965\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>基于大语言模型的智能体评估综述\u003C\u002Fb>\u003C\u002Fi>, 匿名作者等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2503.16416\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.03-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>个性化大型语言模型综述：进展与未来方向\u003C\u002Fb>\u003C\u002Fi>, 匿名作者等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.11528\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.02-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>代理式检索增强生成综述\u003C\u002Fb>\u003C\u002Fi>, 匿名作者等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.09136\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.01-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>基于图的检索增强生成（GraphRAG）\u003C\u002Fb>\u003C\u002Fi>, 匿名作者等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.00309\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.12-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FGraph-RAG\u002FGraphRAG\u002F\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FGraph-RAG\u002FGraphRAG.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>面向大语言模型的代理式强化学习现状：综述\u003C\u002Fb>\u003C\u002Fi>, Zhang 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2509.02547\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.09-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fxhyumiracle\u002FAwesome-AgenticLLM-RL-Papers\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fxhyumiracle\u002FAwesome-AgenticLLM-RL-Papers.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>基准测试\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>评估大语言模型智能体的超长期对话记忆（LOCOMO）\u003C\u002Fb>\u003C\u002Fi>, 匿名作者等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.17753\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FACL-2024.02-blue\" alt=\"ACL徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fsnap-research.github.io\u002Flocomo\u002F\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fsnap-research\u002Flocomo.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>通过增量多轮交互评估大语言模型智能体的记忆\u003C\u002Fb>\u003C\u002Fi>, Hu 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.05257\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.07-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FHUST-AI-HYZ\u002FMemoryAgentBench\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FHUST-AI-HYZ\u002FMemoryAgentBench.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n      \u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fdatasets\u002Fai-hyz\u002FMemoryAgentBench\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fhuggingface.co\u002Fdatasets\u002Fhuggingface\u002Fbadges\u002Fresolve\u002Fmain\u002Fdataset-on-hf-sm.svg\" alt=\"HF数据集\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>大型语言模型的情景记忆生成与评估基准\u003C\u002Fb>\u003C\u002Fi>, 匿名作者等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.13121\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.01-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>论大语言模型智能体的结构化记忆\u003C\u002Fb>\u003C\u002Fi>, 匿名作者等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.15266\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.12-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>HotpotQA：用于多样化、可解释的多跳问答的数据集\u003C\u002Fb>\u003C\u002Fi>, Yang 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F1809.09600\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FEMNLP-2018.09-blue\" alt=\"EMNLP徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fhotpotqa.github.io\u002F\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fhotpotqa\u002Fhotpot.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cb>神经记忆架构\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>神经图灵机\u003C\u002Fb>\u003C\u002Fi>, Graves 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F1410.5401\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2014.10-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>可微分神经计算机\u003C\u002Fb>\u003C\u002Fi>, Graves 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F1610.06258\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2016.10-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fgoogle-deepmind\u002Fdnc\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fgoogle-deepmind\u002Fdnc.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>一种基于脑启发记忆变换的可微分神经计算机\u003C\u002Fb>\u003C\u002Fi>, 匿名作者等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2301.02809\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.01-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>带有记忆恶魔的可微分神经计算机\u003C\u002Fb>\u003C\u002Fi>, 匿名作者等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2211.02987\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.11-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cb>记忆增强型 Transformer\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>记忆化 Transformer\u003C\u002Fb>\u003C\u002Fi>, Wu 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2203.08913\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.03-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>循环记忆 Transformer\u003C\u002Fb>\u003C\u002Fi>, Bulatov 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2207.06881\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNeurIPS-2022.07-blue\" alt=\"NeurIPS Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fbooydar\u002Frecurrent-memory-transformer\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fbooydar\u002Frecurrent-memory-transformer.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>不留下任何上下文：基于 Infini-attention 的高效无限上下文 Transformer\u003C\u002Fb>\u003C\u002Fi>, Munkhdalai 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.07143\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Memformer：用于序列建模的记忆增强型 Transformer\u003C\u002Fb>\u003C\u002Fi>, Wu 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2010.06891\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2020.10-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>标记图灵机\u003C\u002Fb>\u003C\u002Fi>, Ryoo 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2211.09119\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.11-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>TransformerFAM：反馈注意力即工作记忆\u003C\u002Fb>\u003C\u002Fi>, Irie 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.09173\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>生产级记忆系统\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>MemGPT：迈向将大语言模型作为操作系统\u003C\u002Fb>\u003C\u002Fi>, Packer 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.08560\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fresearch.memgpt.ai\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fletta-ai\u002Fletta.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>MemoryBank：用长期记忆增强大型语言模型\u003C\u002Fb>\u003C\u002Fi>, Zhong 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2305.10250\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fzhongwanjun\u002Fmemorybank-siliconfriend\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fzhongwanjun\u002Fmemorybank-siliconfriend.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>MEM0：构建具有可扩展长期记忆的生产就绪型 AI 代理\u003C\u002Fb>\u003C\u002Fi>, Taranjeet 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2504.19413\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fmem0.ai\u002Fresearch\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmem0ai\u002Fmem0.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>MEM1：学习协同记忆与推理以打造高效的长时程智能体\u003C\u002Fb>\u003C\u002Fi>, 匿名作者等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2506.15841\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.06-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmannaandpoem\u002Fopenmanus\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmannaandpoem\u002Fopenmanus.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>A-MEM：面向 LLM 代理的主体式记忆\u003C\u002Fb>\u003C\u002Fi>, 匿名作者等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.12110\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fagiresearch\u002FA-mem\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fagiresearch\u002FA-mem.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>MemAgent：利用多卷积强化学习驱动的记忆智能体重塑长上下文 LLM\u003C\u002Fb>\u003C\u002Fi>, 匿名作者等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.02259\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.07-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>AI 代理的记忆操作系统\u003C\u002Fb>\u003C\u002Fi>, Kang 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2506.06326\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FBAI-LAB\u002FMemoryOS\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FBAI-LAB\u002FMemoryOS.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cb>基于图的记忆系统\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>arigraph: 使用情景记忆为 LLM 代理学习知识图谱世界模型\u003C\u002Fb>\u003C\u002Fi>, 匿名等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.04363\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.07-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Zep：用于代理记忆的时序知识图谱架构\u003C\u002Fb>\u003C\u002Fi>, 匿名等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.13956\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.01-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fgetzep\u002Fgraphiti\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fgetzep\u002Fgraphiti.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>KG-Agent：面向知识图谱复杂推理的高效自主代理框架\u003C\u002Fb>\u003C\u002Fi>, 匿名等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.11163\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>GraphReader：构建基于图的代理以增强大型语言模型的长上下文能力\u003C\u002Fb>\u003C\u002Fi>, 匿名等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2406.14550\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.06-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>从局部到全局：一种面向查询聚焦摘要的 GraphRAG 方法\u003C\u002Fb>\u003C\u002Fi>, Edge 等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.16130\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002Fgraphrag\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmicrosoft\u002Fgraphrag.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>知识图谱引导的检索增强生成\u003C\u002Fb>\u003C\u002Fi>, Zhu 等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.06864\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cb>情景记忆与工作记忆\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Larimar：具有情景记忆控制的大语言模型\u003C\u002Fb>\u003C\u002Fi>, Goyal 等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2403.11901\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICML-2024.03-blue\" alt=\"ICML Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>EM-LLM：适用于无限上下文 LLM 的类人情景记忆\u003C\u002Fb>\u003C\u002Fi>, 匿名等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.09450\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2024.07-blue\" alt=\"ICLR Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fem-llm\u002FEM-LLM-model\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fem-llm\u002FEM-LLM-model.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>具有可控工作记忆的大语言模型\u003C\u002Fb>\u003C\u002Fi>, Goyal 等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2211.05110\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.11-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>赋能大语言模型代理的工作记忆\u003C\u002Fb>\u003C\u002Fi>, 匿名等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.17259\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.12-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cb>对话记忆\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>MemoChat：调优 LLM 以利用备忘录实现持续的长程开放域对话\u003C\u002Fb>\u003C\u002Fi>, 匿名等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2308.08239\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.08-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>思维在记忆中：回忆与后思使 LLM 具备长期记忆\u003C\u002Fb>\u003C\u002Fi>, 匿名等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2311.08719\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.11-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>生成式代理：人类行为的交互式模拟物\u003C\u002Fb>\u003C\u002Fi>, Park 等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2304.03442\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>大型语言模型的自控记忆框架\u003C\u002Fb>\u003C\u002Fi>, 匿名等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2304.13343\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cb>来自主要会议的基础性综述论文\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>AUTOPROMPT：通过自动生成提示从语言模型中提取知识\u003C\u002Fb>\u003C\u002Fi>, Shin 等, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FEMNLP-2020-blue\" alt=\"EMNLP Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fucinlp\u002Fautoprompt\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fucinlp\u002Fautoprompt.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>参数高效提示调优中的规模效应\u003C\u002Fb>\u003C\u002Fi>, Lester 等, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FEMNLP-2021-blue\" alt=\"EMNLP Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fgoogle-research\u002Fprompt-tuning\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fgoogle-research\u002Fprompt-tuning.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>前缀调优：优化连续提示以进行生成\u003C\u002Fb>\u003C\u002Fi>, Li 等, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FACL-2021-blue\" alt=\"ACL Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FXiangLi1999\u002FPrefixTuning\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FXiangLi1999\u002FPrefixTuning.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>上下文学习作为隐式贝叶斯推断的解释\u003C\u002Fb>\u003C\u002Fi>, Xie 等, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2022-blue\" alt=\"ICLR Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fp-lambda\u002Fincontext-learning\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fp-lambda\u002Fincontext-learning.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>重新思考演示的作用：是什么让上下文学习奏效？\u003C\u002Fb>\u003C\u002Fi>, Min 等, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FEMNLP-2022-blue\" alt=\"EMNLP Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FAlrope123\u002Frethinking-demonstrations\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FAlrope123\u002Frethinking-demonstrations.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>额外的 RAG 和检索综述\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>用于 AI 生成内容的检索增强生成：综述\u003C\u002Fb>\u003C\u002Fi>, 各方, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.19473\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FPKU-DAIR\u002FRAG-Survey\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FPKU-DAIR\u002FRAG-Survey.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>检索增强生成 (RAG) 及其扩展：关于如何更明智地让您的大语言模型使用外部数据的综合综述\u003C\u002Fb>\u003C\u002Fi>, 各方, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2409.14924\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.09-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>大型语言模型 (LLMs)：综述、技术框架与未来挑战\u003C\u002Fb>\u003C\u002Fi>, 各方, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAIR-2024-blue\" alt=\"AIR Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\n---\n\n\n\n## 🏗️ 上下文工程的定义\n\n> **上下文不仅仅是用户发送给大语言模型的单个提示。上下文是在推理时提供给大语言模型的完整信息载荷，包含了模型为合理完成给定任务所需的所有结构化信息组件。**\n\n### 大语言模型生成\n\n要正式定义上下文工程，我们首先需要从数学上描述大语言模型的生成过程。让我们将大语言模型建模为一个概率函数：\n\n$$P(\\text{output} | \\text{context}) = \\prod_{t=1}^T P(\\text{token}_t | \\text{previous tokens}, \\text{context})$$\n\n其中：\n- $\\text{context}$ 表示提供给大语言模型的完整输入信息\n- $\\text{output}$ 表示生成的响应序列\n- $P(\\text{token}_t | \\text{previous tokens}, \\text{context})$ 是在给定上下文的情况下生成每个 token 的概率\n\n### 上下文的定义\n\n在传统的提示工程中，上下文被视为一个简单的字符串：\n$$\\text{context} = \\text{prompt}$$\n\n然而，在上下文工程中，我们将上下文分解为多个结构化的组成部分：\n\n$$\\text{context} = \\text{Assemble}(\\text{instructions}, \\text{knowledge}, \\text{tools}, \\text{memory}, \\text{state}, \\text{query})$$\n\n其中 $\\text{Assemble}$ 是一个上下文组装函数，负责协调：\n- $\\text{instructions}$：系统提示和规则\n- $\\text{knowledge}$：检索到的相关信息\n- $\\text{tools}$：可用的函数定义\n- $\\text{memory}$：对话历史和已学习的事实\n- $\\text{state}$：当前的世界\u002F用户状态\n- $\\text{query}$：用户的即时请求\n\n### 上下文工程的定义\n\n**上下文工程** 被正式定义为以下优化问题：\n\n$$\\text{Assemble}^* = \\arg\\max_{\\text{Assemble}} \\mathbb{E} [\\text{Reward}(\\text{LLM}(\\text{context}), \\text{target})]$$\n\n受以下约束：\n- $|\\text{context}| \\leq \\text{MaxTokens} \\text{(上下文窗口限制)}$\n- $\\text{knowledge} = \\text{Retrieve}(\\text{query}, \\text{database})$\n- $\\text{memory} = \\text{Select}(\\text{history}, \\text{query})$\n- $\\text{state} = \\text{Extract}(\\text{world})$\n\n其中：\n- $\\text{Reward}$ 衡量生成响应的质量\n- $\\text{Retrieve}$、$\\text{Select}$、$\\text{Extract}$ 是用于信息收集的函数\n\n### 动态上下文编排\n\n上下文的组装可以分解为：\n\n$$\\text{context} = \\text{Concat}(\\text{Format}(\\text{instructions}), \\text{Format}(\\text{knowledge}), \\text{Format}(\\text{tools}), \\text{Format}(\\text{memory}), \\text{Format}(\\text{query}))$$\n\n其中 $\\text{Format}$ 表示特定组件的结构化处理，而 $\\text{Concat}$ 则按照令牌限制和最佳位置将其组装起来。\n\n因此，**上下文工程** 就是设计和优化这些组装与格式化函数，以最大限度地提高任务性能的学科。\n\n### 数学原理\n\n从这一形式化中，我们得出四个基本原则：\n\n1. **系统级优化**：上下文生成是一个关于组装函数的多目标优化问题，而不是简单的字符串操作。\n\n2. **动态适应**：上下文组装函数会根据每次推理时的 $\\text{query}$ 和 $\\text{state}$ 进行调整：$\\text{Assemble}(\\cdot | \\text{query}, \\text{state})$。\n\n3. **信息论最优性**：检索函数会最大化相关信息：$\\text{Retrieve} = \\arg\\max \\text{Relevance}(\\text{knowledge}, \\text{query})$。\n\n4. **结构敏感性**：格式化函数编码了与大语言模型处理能力相匹配的结构。\n\n### 理论框架：贝叶斯上下文推断\n\n上下文工程可以在贝叶斯框架内进行形式化，其中最优的上下文被推断出来：\n\n$$P(\\text{context} | \\text{query}, \\text{history}, \\text{world}) \\propto P(\\text{query} | \\text{context}) \\cdot P(\\text{context} | \\text{history}, \\text{world})$$\n\n其中：\n- $P(\\text{query} | \\text{context})$ 建模查询与上下文的兼容性\n- $P(\\text{context} | \\text{history}, \\text{world})$ 表示先验上下文概率\n\n最优的上下文组装变为：\n\n$$\\text{context}^* = \\arg\\max_{\\text{context}} P(\\text{answer} | \\text{query}, \\text{context}) \\cdot P(\\text{context} | \\text{query}, \\text{history}, \\text{world})$$\n\n这种贝叶斯公式化能够实现：\n- **不确定性量化**：对上下文相关性的置信度建模\n- **自适应检索**：根据反馈更新上下文信念\n- **多步推理**：在交互过程中保持上下文分布\n\n### 对比\n\n| 维度         | 提示工程                     | 上下文工程                     |\n|--------------|------------------------------|--------------------------------|\n| **数学模型** | $\\text{context} = \\text{prompt}$（静态） | $\\text{context} = \\text{Assemble}(...)$（动态） |\n| **优化目标** | $\\arg\\max_{\\text{prompt}} P(\\text{answer} \\mid \\text{query}, \\text{prompt})$ | $\\arg\\max_{\\text{Assemble}} \\mathbb{E}[\\text{Reward}(...)]$ |\n| **复杂度**   | $O(1)$ 上下文组装            | $O(n)$ 多组件优化              |\n| **信息理论** | 固定的信息含量               | 自适应的信息最大化             |\n| **状态管理** | 无状态函数                   | 有状态，带有 $\\text{memory}(\\text{history}, \\text{query})$ |\n| **可扩展性** | 与提示长度线性相关           | 通过压缩\u002F过滤呈亚线性          |\n| **错误分析** | 手动检查提示                 | 系统评估组装组件               |\n\n\n\n---\n\n## 🌐 相关博客\n\n- [“上下文工程”的兴起](https:\u002F\u002Fblog.langchain.com\u002Fthe-rise-of-context-engineering\u002F)\n- [AI领域的新技能不是提示工程，而是上下文工程](https:\u002F\u002Fwww.philschmid.de\u002Fcontext-engineering)\n- [davidkimai\u002FContext-Engineering: “上下文工程是一门精妙的艺术与科学，旨在为下一步操作恰到好处地填充上下文窗口。”](https:\u002F\u002Fgithub.com\u002Fdavidkimai\u002FContext-Engineering)\n- [上下文工程是AI智能体的运行时 | 作者：Bijit Ghosh | 2025年6月 | Medium](https:\u002F\u002Fmedium.com\u002F@bijit211987\u002Fcontext-engineering-is-runtime-of-ai-agents-411c9b2ef1cb)\n- [上下文工程](https:\u002F\u002Fblog.langchain.com\u002Fcontext-engineering-for-agents\u002F)\n- [面向智能体的上下文工程](https:\u002F\u002Frlancemartin.github.io\u002F2025\u002F06\u002F23\u002Fcontext_engineering\u002F)\n- [Cognition | 不要构建多智能体](https:\u002F\u002Fcognition.ai\u002Fblog\u002Fdont-build-multi-agents)\n- [从Prompt Engineering到Context Engineering - 53AI-AI知识库|大模型知识库|大模型训练|智能体开发](https:\u002F\u002Fwww.53ai.com\u002Fnews\u002Ftishicikuangjia\u002F2025062727685.html)\n\n### 社交媒体与演讲\n\n- [30分钟掌握Claude代码](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=6eBSHbLKuN0)\n- [面向智能体的上下文工程](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=4GiqzUHD5AA)\n- [Andrej Karpathy在X平台发文：“支持用‘上下文工程’取代‘提示工程’”](https:\u002F\u002Fx.com\u002Fkarpathy\u002Fstatus\u002F1937902205765607626?ref=blog.langchain.com)\n- [复旦大学\u002F上海创智学院邱锡鹏：上下文扩展，通往AGI的下一幕](https:\u002F\u002Fmp.weixin.qq.com\u002Fs\u002FKnej0qbyr5j5KX_BO7FGew)\n\n---\n\n## 🤔 为什么需要上下文工程？\n\n### 范式转变：从战术到战略\n\n从提示工程到上下文工程的演进，标志着AI系统设计的根本性成熟。正如Andrej Karpathy、Tobi Lutke和Simon Willison等重要人物所指出的那样，“提示工程”这一术语已被泛化为仅仅意味着“向聊天机器人输入内容”，而无法体现工业级LLM应用所需的复杂性。\n\n### 1. 当前方法的基本挑战\n\n#### 人类意图表达的挑战\n- **人类意图不明确**：人类在用自然语言表达意图时，常常模糊、不完整或含糊不清。\n- **AI对人类意图理解不足**：AI系统难以完全理解复杂的意图，尤其是涉及隐含背景或文化差异的情况。\n- **AI过度字面解读**：AI常会过于字面地理解人类指令，从而忽略其潜在意图或语境意义。\n\n#### 复杂知识需求\n单个模型本身无法解决需要以下条件的复杂问题：\n- **(1) 大规模外部知识**：超出模型容量的海量外部知识。\n- **(2) 准确的外部知识**：模型可能不具备的精确、最新信息。\n- **(3) 新兴的外部知识**：模型训练完成后出现的新知识。\n\n**静态知识的局限性：**\n- **静态知识问题**：预训练模型包含的知识是静态的，会逐渐过时。\n- **知识截止期**：模型无法访问超出其训练数据范围的信息。\n- **领域特定的知识缺口**：模型缺乏针对特定行业或应用场景的专业知识。\n\n#### 可靠性和可信度问题\n- **AI幻觉**：当缺乏适当上下文时，LLM会生成看似合理但事实错误的信息。\n- **来源不可追溯**：生成信息缺乏明确的来源标注。\n- **置信度校准不足**：即使生成错误信息，模型仍可能表现得非常自信。\n- **透明度不足**：无法追踪结论是如何得出的。\n- **责任归属问题**：难以验证AI生成内容的可靠性。\n\n### 2. 静态提示的局限性\n\n#### 从字符串到系统\n传统提示将上下文视为静态字符串，而企业级应用则需要：\n- **动态信息组装**：根据具体用户和查询实时构建的上下文。\n- **多源整合**：结合数据库、API、文档和实时数据。\n- **状态管理**：维护对话历史、用户偏好和工作流状态。\n- **工具编排**：协调外部函数调用和API交互。\n\n#### “电影制作”类比\n如果提示工程只是为演员写一句台词，那么上下文工程则是搭建整个场景、设计灯光、提供详尽的背景故事并指导整场戏的过程。只有在丰富且精心构建的环境中，这句台词才能发挥出预期的效果。\n\n### 3. 企业级与生产环境的需求\n\n#### 上下文失败成为新的瓶颈\n现代智能体系统中的大多数故障已不再归因于核心模型的推理能力，而是源于“上下文失败”。真正的工程挑战不在于该提出什么问题，而在于如何确保模型拥有所有必要的背景信息、数据、工具和记忆，以便有意义且可靠地作出回答。\n\n#### 超越简单任务的可扩展性\n虽然提示工程足以应对简单的独立任务，但在扩展到以下场景时便会失效：\n- **复杂的多步骤应用**\n- **数据密集型企业环境**\n- **有状态的长期运行工作流**\n- **多用户、多租户系统**\n\n#### 可靠性与一致性\n企业级应用要求：\n- **确定性行为**：在不同上下文和用户之间保持可预测的输出。\n- **错误处理**：当信息不完整或相互矛盾时，能够优雅降级。\n- **审计追踪**：透明地记录上下文如何影响模型决策。\n- **合规性**：满足数据处理和决策制定方面的监管要求。\n\n#### 经济与运营效率\n上下文工程能够实现：\n- **成本优化**：在RAG和长上下文方法之间做出战略性选择。\n- **延迟管理**：高效检索信息并组装上下文。\n- **资源利用**：优化有限上下文窗口和计算资源的使用。\n- **维护可扩展性**：系统化的方法用于更新和管理知识库。\n\n上下文工程为管理状态、整合多样化的数据源以及在这些严苛场景中保持一致性提供了架构基础。\n\n### 4. 认知与信息科学基础\n\n#### 人工具身\n大语言模型本质上是“缸中之脑”——强大的推理引擎，却缺乏与具体环境的连接。上下文工程提供了：\n- **合成感知系统**：将检索机制作为人工感知\n- **代理具身**：将工具使用视为人工行动能力\n- **人工记忆**：结构化信息的存储与检索\n\n#### 大规模信息检索\n上下文工程解决了信息检索中的根本挑战，即“用户”并非人类，而是一个AI智能体。这需要：\n- **语义理解**：弥合意图与表达之间的鸿沟\n- **相关性优化**：对海量知识库进行排序与过滤\n- **查询转换**：将模糊请求转化为精确的检索操作\n\n### 5. AI系统架构的未来\n上下文工程将AI开发从一系列“提示技巧”提升为一门严谨的系统架构学科。它将操作系统设计、内存管理及分布式系统领域的数十年经验应用于基于大语言模型的应用所面临的独特挑战。\n\n这一学科为在生产环境中充分发挥大语言模型的潜力奠定了基础，推动了从一次性文本生成向能够在复杂、动态环境中可靠运行的自主智能体和高级AI助手的转变。\n\n---\n\n## 🔧 组件、技术和架构\n\n### 上下文扩展\n\n\u003Cb>位置插值与扩展技术\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>通过位置插值扩展大型语言模型的上下文窗口\u003C\u002Fb>\u003C\u002Fi>, Chen等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2306.15595\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.06-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FMath1019\u002FExtend_Context_Window_Position_Interpolation\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FMath1019\u002FExtend_Context_Window_Position_Interpolation.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>YaRN：高效扩展大型语言模型的上下文窗口\u003C\u002Fb>\u003C\u002Fi>, Peng等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2309.00071\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2024.01-blue\" alt=\"ICLR徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fjquesnelle\u002Fyarn\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fjquesnelle\u002Fyarn.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LongRoPE：将LLM上下文窗口扩展至超过200万标记\u003C\u002Fb>\u003C\u002Fi>, Ding等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.13753\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICML-2024.02-blue\" alt=\"ICML徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002FLongRoPE\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmicrosoft\u002FLongRoPE.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LongRoPE2：近乎无损的大语言模型上下文窗口扩展\u003C\u002Fb>\u003C\u002Fi>, Shang等，\u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICML-2025.05-blue\" alt=\"ICML徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002FLongRoPE\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmicrosoft\u002FLongRoPE.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>内存高效的注意力机制\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>快速多极注意力：一种用于长序列的分治型注意力机制\u003C\u002Fb>\u003C\u002Fi>, Kang等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.11960\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2024.02-blue\" alt=\"ICLR徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fyanmingk\u002FFMA\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fyanmingk\u002FFMA.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>不遗漏任何上下文：采用Infini-attention的高效无限上下文Transformer\u003C\u002Fb>\u003C\u002Fi>, Munkhdalai等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.07143\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fjlamprou\u002FInfini-Attention\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fjlamprou\u002FInfini-Attention.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>DuoAttention：结合检索与流式处理头的高效长上下文LLM推理\u003C\u002Fb>\u003C\u002Fi>, Xiao等，\u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2025.01-blue\" alt=\"ICLR徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmit-han-lab\u002Fduo-attention\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmit-han-lab\u002Fduo-attention.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Star Attention：高效处理长序列的LLM推理\u003C\u002Fb>\u003C\u002Fi>, Acharya等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2411.17116\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.11-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FStar-Attention\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FNVIDIA\u002FStar-Attention.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>超长序列处理（10万+标记）\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>TokenSwift：无损加速超长序列生成\u003C\u002Fb>\u003C\u002Fi>, Wu等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.18890\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICML-2025.02-blue\" alt=\"ICML徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fbigai-nlco\u002FTokenSwift\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fbigai-nlco\u002FTokenSwift.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LongHeads：多头注意力其实是一种长上下文处理器\u003C\u002Fb>\u003C\u002Fi>, Lu等，\u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FEMNLP-2024.11-blue\" alt=\"EMNLP徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FLuLuLuyi\u002FLongHeads\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FLuLuLuyi\u002FLongHeads.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>∞Bench：将长上下文评估扩展至10万标记以上\u003C\u002Fb>\u003C\u002Fi>, Bai等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.00359\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FACL-2024.06-blue\" alt=\"ACL徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FOpenBMB\u002FInfiniteBench\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FOpenBMB\u002FInfiniteBench.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>上下文长度扩展的综述与方法\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>超越极限：大型语言模型上下文长度扩展技术综述\u003C\u002Fb>\u003C\u002Fi>, 各方, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.02244\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.02-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LLM中长上下文扩展与泛化能力的受控研究\u003C\u002Fb>\u003C\u002Fi>, 各方, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2409.12181\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.09-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FLeooyii\u002FLCEG\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FLeooyii\u002FLCEG.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>选择性注意力：通过原则性的上下文控制提升Transformer性能\u003C\u002Fb>\u003C\u002Fi>, 各方, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNeurIPS-2024-blue\" alt=\"NeurIPS徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fumich-sota\u002Fselective_attention\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fumich-sota\u002Fselective_attention.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cb>具备复杂上下文理解能力的视觉-语言模型\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>迈向以LLM为中心的多模态融合：集成策略与技术综述\u003C\u002Fb>\u003C\u002Fi>, An等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2506.04788\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.01-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>浏览与聚焦：通过先验-LLM上下文融合理解多模态内容\u003C\u002Fb>\u003C\u002Fi>, Wang等, \u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.18653\u002Fv1\u002F2024.acl-long.605\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FACL-2024.08-blue\" alt=\"ACL徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FTHUNLP-MT\u002FBrote\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FTHUNLP-MT\u002FBrote.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>V2PE：利用可变视觉位置编码提升视觉-语言模型的多模态长上下文能力\u003C\u002Fb>\u003C\u002Fi>, Dai等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.09616\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.12-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FOpenGVLab\u002FV2PE\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FOpenGVLab\u002FV2PE.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Flamingo：用于少样本学习的视觉语言模型\u003C\u002Fb>\u003C\u002Fi>, Alayrac等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2204.14198\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNeurIPS-2022.04-blue\" alt=\"NeurIPS徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Flucidrains\u002Fflamingo-pytorch\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Flucidrains\u002Fflamingo-pytorch.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n\u003Cb>视听上下文融合与处理\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>对齐更好，聆听更好：面向视听大型语言模型\u003C\u002Fb>\u003C\u002Fi>, Guo等, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2025.01-blue\" alt=\"ICLR徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>AVicuna：具有交错器和上下文边界对齐功能的视听LLM，用于时序参照对话\u003C\u002Fb>\u003C\u002Fi>, Chen等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2403.16276\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.03-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>SonicVisionLM：用视觉语言模型播放声音\u003C\u002Fb>\u003C\u002Fi>, Xie等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.04394\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FCVPR-2024.01-blue\" alt=\"CVPR徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FYusiissy\u002FSonicVisionLM\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FYusiissy\u002FSonicVisionLM.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>SAVEn-Vid：协同视听融合，增强长视频上下文理解能力\u003C\u002Fb>\u003C\u002Fi>, Li等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2411.16213\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.11-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FLJungang\u002FSAVEn-Vid\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FLJungang\u002FSAVEn-Vid.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n\n\u003Cb>多模态提示工程与上下文设计\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>CaMML：面向大型模型的上下文感知多模态学习者\u003C\u002Fb>\u003C\u002Fi>, Chen等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.11406\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FACL-2024.08-blue\" alt=\"ACL徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>大型视觉-语言模型的视觉内上下文学习\u003C\u002Fb>\u003C\u002Fi>, Zhou等, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FACL-2024.08-blue\" alt=\"ACL徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>CAMA：利用上下文感知调制注意力增强多模态内上下文学习\u003C\u002Fb>\u003C\u002Fi>, Li等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2505.17097\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.05-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n\n\u003Cb>CVPR 2024视觉-语言领域进展\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>CogAgent：用于GUI代理的视觉语言模型\u003C\u002Fb>\u003C\u002Fi>, 各方, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FCVPR-2024-blue\" alt=\"CVPR徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FTHUDM\u002FCogAgent\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FTHUDM\u002FCogAgent.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LISA：基于大型语言模型的推理分割\u003C\u002Fb>\u003C\u002Fi>, 各方, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FCVPR-2024-blue\" alt=\"CVPR徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fdvlab-research\u002FLISA\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fdvlab-research\u002FLISA.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>对比语言-图像学习的可重复缩放规律\u003C\u002Fb>\u003C\u002Fi>, 各方, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FCVPR-2023-blue\" alt=\"CVPR徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FLAION-AI\u002Fscaling-laws-openclip\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FLAION-AI\u002Fscaling-laws-openclip.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n\n\u003Cb>视频与时序理解\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>大型语言模型在视频理解中的应用：综述\u003C\u002Fb>\u003C\u002Fi>, 各方, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.17432\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.12-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fyunlong10\u002FAwesome-LLMs-for-Video-Understanding\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fyunlong10\u002FAwesome-LLMs-for-Video-Understanding.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n### 生产环境中的上下文管理\n\n在智能体时代，上下文工程越来越意味着**运行时上下文管理**，而不仅仅是提示词的构建。如今，生产系统依赖于上下文压缩、缓存、基于工件的状态管理以及作用域限定的指令加载，以确保长周期智能体的高效性和可控性。\n\n\u003Cb>运行时上下文管理模式\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>OpenAI 智能体指南\u003C\u002Fb>\u003C\u002Fi>, OpenAI，\u003Ca href=\"https:\u002F\u002Fplatform.openai.com\u002Fdocs\u002Fguides\u002Fagents\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpenAI-2026-blue\" alt=\"OpenAI 标志\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>OpenAI 工具：对话状态、提示缓存与压缩\u003C\u002Fb>\u003C\u002Fi>, OpenAI，\u003Ca href=\"https:\u002F\u002Fdevelopers.openai.com\u002Fapi\u002Fdocs\u002Fguides\u002Ftools\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpenAI-2026-blue\" alt=\"OpenAI 标志\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Google ADK：上下文缓存与上下文压缩\u003C\u002Fb>\u003C\u002Fi>, Google，\u003Ca href=\"https:\u002F\u002Fgoogle.github.io\u002Fadk-docs\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FGoogle-2026-blue\" alt=\"Google 标志\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Claude Code 内存与作用域限定的项目指令\u003C\u002Fb>\u003C\u002Fi>, Anthropic，\u003Ca href=\"https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fdocs\u002Fclaude-code\u002Fmemory\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAnthropic-2026-blue\" alt=\"Anthropic 标志\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LangChain Deep Agents：基于文件系统的上下文管理\u003C\u002Fb>\u003C\u002Fi>, LangChain，\u003Ca href=\"https:\u002F\u002Fdocs.langchain.com\u002Foss\u002Fpython\u002Fdeepagents\u002Foverview\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLangChain-2026-blue\" alt=\"LangChain 标志\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>生产设计问题\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>何时应将状态保留在提示中，而非移至文件、内存存储或外部工具？\u003C\u002Fb>\u003C\u002Fi>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>如何在不丢失出处信息、指令或活动计划的情况下压缩长时间运行的会话？\u003C\u002Fb>\u003C\u002Fi>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>如何根据路径、任务或子智能体有条件地加载项目规则，而不是全局加载？\u003C\u002Fb>\u003C\u002Fi>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>如何将提示缓存与内存写入及检索的新鲜度相结合？\u003C\u002Fb>\u003C\u002Fi>\u003C\u002Fli>\n\u003C\u002Ful>\n\n### 结构化数据集成\n\n\u003Cb>知识图谱增强型语言模型\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>共同学习：面向下游任务的预训练知识图谱增强型大语言模型联合多任务微调\u003C\u002Fb>\u003C\u002Fi>, Martynova 等人，\u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.18653\u002Fv1\u002F2025.genaik-1.2\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICCL-2025.01-blue\" alt=\"ICCL徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FVloods\u002Fmultitask_finetune\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FVloods\u002Fmultitask_finetune.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>知识图谱调优：基于人类反馈的实时大型语言模型个性化\u003C\u002Fb>\u003C\u002Fi>, Sun 等人，\u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2025.02-blue\" alt=\"ICLR徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>知识图谱引导的检索增强生成\u003C\u002Fb>\u003C\u002Fi>, Zhu 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.06864\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.02-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fnju-websoft\u002FKG2RAG\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstarsnju-websoft\u002FKG2RAG.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>KGLA：用于客户服务的知识图谱增强型语言代理\u003C\u002Fb>\u003C\u002Fi>, 匿名作者等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.19627\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.10-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\u003Cb>图神经网络与语言模型结合\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>大型语言模型是上下文图学习者吗？\u003C\u002Fb>\u003C\u002Fi>, Li 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.13562\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.02-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fyunlong10\u002FAwesome-LLMs-for-Video-Understanding\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fyunlong10\u002FAwesome-LLMs-for-Video-Understanding.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>让我们问问GNN：赋能大型语言模型进行图上下文学习\u003C\u002Fb>\u003C\u002Fi>, Hu 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.07074\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FEMNLP-2024.11-blue\" alt=\"EMNLP徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fppsmk388\u002FAskGNN\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fppsmk388\u002FAskGNN.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>GL-Fusion：重新思考图神经网络与大型语言模型的结合\u003C\u002Fb>\u003C\u002Fi>, Yang 等人，\u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2025.02-blue\" alt=\"ICLR徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>NT-LLM：一种将图结构整合到大型语言模型中的新型节点标记器\u003C\u002Fb>\u003C\u002Fi>, Ji 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.10743\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.10-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\u003Cb>结构化数据集成\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>CoddLLM：赋能大型语言模型进行数据分析\u003C\u002Fb>\u003C\u002Fi>, 作者等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.00329\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.02-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>结构引导的大型语言模型用于文本到SQL生成\u003C\u002Fb>\u003C\u002Fi>, 作者等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.13284\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.02-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>StructuredRAG：利用大型语言模型进行JSON响应格式化\u003C\u002Fb>\u003C\u002Fi>, 作者等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2408.11061\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.08-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fweaviate\u002Fstructured-rag\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fweaviate\u002Fstructured-rag.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>基础性知识图谱-语言模型融合方法\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>统一大型语言模型和知识图谱：路线图\u003C\u002Fb>\u003C\u002Fi>, 各方，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2306.08302\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.06-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FRManLuo\u002FAwesome-LLM-KG?tab=readme-ov-file\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FRManLuo\u002FAwesome-LLM-KG?tab=readme-ov-file.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>结合知识图谱和大型语言模型\u003C\u002Fb>\u003C\u002Fi>, 各方，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.06564\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.07-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>众人对一人：高效整合大型语言模型以实现图神经网络中的消息传递\u003C\u002Fb>\u003C\u002Fi>, 各方，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.14996\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.07-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>大型语言模型用于图学习\u003C\u002Fb>\u003C\u002Fi>, 各方，\u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FWWW-2024-blue\" alt=\"WWW徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n### 自生成上下文\n\n\u003Cb>自监督上下文生成与增强\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>SelfCite：用于大型语言模型中上下文归因的自监督对齐\u003C\u002Fb>\u003C\u002Fi>, Chuang 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.09604\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Ffacebookresearch\u002FSelfCite\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Ffacebookresearch\u002FSelfCite.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>自监督提示优化\u003C\u002Fb>\u003C\u002Fi>, Xiang 等人，\u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FCoRR-2025.01-orange\" alt=\"CoRR Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FFoundationAgents\u002FMetaGPT\u002Ftree\u002Fmain\u002Fexamples\u002Fspo\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FFoundationAgents\u002FMetaGPT\u002Ftree\u002Fmain\u002Fexamples\u002Fspo.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>SCOPE：一种用于提升条件文本生成忠实性的自监督框架\u003C\u002Fb>\u003C\u002Fi>, Duong 等人，\u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2025.01-blue\" alt=\"ICLR Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fsngdng\u002Fscope-faithfulness\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fsngdng\u002Fscope-faithfulness.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>能够自动生成上下文的推理模型\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>自我一致性提升语言模型中的思维链推理\u003C\u002Fb>\u003C\u002Fi>, Wang 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2203.11171\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2023.02-blue\" alt=\"ICLR Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>思维之树：利用大型语言模型进行深思熟虑的问题解决\u003C\u002Fb>\u003C\u002Fi>, Yao 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2305.10601\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fprinceton-nlp\u002Ftree-of-thought-llm\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fprinceton-nlp\u002Ftree-of-thought-llm.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>从自训练视角重新思考思维链\u003C\u002Fb>\u003C\u002Fi>, Wu 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.10827\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.12-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fzongqianwu\u002FST-COT\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fzongqianwu\u002FST-COT.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>大型语言模型的自主树搜索能力\u003C\u002Fb>\u003C\u002Fi>, 作者等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.10686\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.10-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FZheyuAqaZhang\u002FAutonomous-Tree-search\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FZheyuAqaZhang\u002FAutonomous-Tree-search.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\u003Cb>迭代式上下文精炼与自我改进\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Self-Refine：基于自我反馈的迭代精炼\u003C\u002Fb>\u003C\u002Fi>, Madaan 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2303.17651\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.03-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmadaan\u002Fself-refine\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmadaan\u002Fself-refine.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>反思、重试、奖励：通过强化学习实现自我改进的语言模型\u003C\u002Fb>\u003C\u002Fi>, 作者等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2505.24726\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>大型语言模型可在长上下文推理中自我改进\u003C\u002Fb>\u003C\u002Fi>, Li 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2411.08147\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.11-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FSihengLi99\u002FSEALONG\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FSihengLi99\u002FSEALONG.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>AlphaCodium代码生成：从提示工程到流程工程\u003C\u002Fb>\u003C\u002Fi>, Oren 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.08500\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa> \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FCodium-ai\u002Falphacodium\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FCodium-ai\u002Falphacodium.svg?style=social\" alt=\"GitHub stars\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>语言智能体树搜索统一了语言模型中的推理、行动与规划\u003C\u002Fb>\u003C\u002Fi>, Zhou 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.04406\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.10-red\" alt=\"arXiv Badge\">\u003C\u002Fa> \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fandyz245\u002FLanguage-Agent-Tree-Search\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fandyz245\u002FLanguage-Agent-Tree-Search.svg?style=social\" alt=\"GitHub stars\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\u003Cb>元学习与自主上下文演化\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>大型语言模型中的元内上下文学习\u003C\u002Fb>\u003C\u002Fi>, Coda-Forno 等人，\u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNeurIPS-2023.12-blue\" alt=\"NeurIPS Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>EvoPrompt：将LLM与进化算法结合可产生强大的提示优化器\u003C\u002Fb>\u003C\u002Fi>, Guo 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2309.08532\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2024.01-blue\" alt=\"ICLR Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fbeeevita\u002FEvoPrompt\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fbeeevita\u002FEvoPrompt.svg?style=social\" alt=\"GitHub stars\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>AutoPDL：面向LLM代理的自动提示优化\u003C\u002Fb>\u003C\u002Fi>, Spiess 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2504.04365\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAutoML-2025.04-orange\" alt=\"AutoML Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Agent-Pro：通过基于提案的编程学习进化编码代理\u003C\u002Fb>\u003C\u002Fi>, Zhang 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.17574\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\u003Cb>思维链研究的基础性成果\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>思维链提示能激发大型语言模型的推理能力\u003C\u002Fb>\u003C\u002Fi>, Wei 等人，\u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNeurIPS-2022-blue\" alt=\"NeurIPS Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\n---\n\n## 🛠️ 实现与挑战\n\n### 0. 智能体框架与运行时系统\n\n到2026年，上下文工程中许多最重要的进展不再仅仅存在于提示词之中。它们已经融入**智能体框架**：即管理计划、子智能体、检查点、文件、审批流程、工具执行以及故障恢复的运行时循环。正是在这里，上下文工程演变为智能体工程。\n\n\u003Cb>框架与运行时设计参考\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>构建高效智能体\u003C\u002Fb>\u003C\u002Fi>, Anthropic, \u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Fbuilding-effective-agents\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAnthropic-2024.12-blue\" alt=\"Anthropic徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>OpenAI智能体指南\u003C\u002Fb>\u003C\u002Fi>, OpenAI, \u003Ca href=\"https:\u002F\u002Fplatform.openai.com\u002Fdocs\u002Fguides\u002Fagents\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpenAI-2026-blue\" alt=\"OpenAI徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Google智能体开发套件（ADK）\u003C\u002Fb>\u003C\u002Fi>, Google, \u003Ca href=\"https:\u002F\u002Fgoogle.github.io\u002Fadk-docs\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FGoogle-2026-blue\" alt=\"Google徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LangChain深度智能体概述\u003C\u002Fb>\u003C\u002Fi>, LangChain, \u003Ca href=\"https:\u002F\u002Fdocs.langchain.com\u002Foss\u002Fpython\u002Fdeepagents\u002Foverview\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLangChain-2026-blue\" alt=\"LangChain徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Microsoft智能体框架概述\u003C\u002Fb>\u003C\u002Fi>, Microsoft, \u003Ca href=\"https:\u002F\u002Flearn.microsoft.com\u002Fen-us\u002Fagent-framework\u002Fuser-guide\u002Foverview\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FMicrosoft-2026-blue\" alt=\"Microsoft徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>核心运行时关注点\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>规划与分解\u003C\u002Fb>\u003C\u002Fi>：如何将任务拆解为可管理的单元\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>持久化执行\u003C\u002Fb>\u003C\u002Fi>：如何对智能体状态进行检查点保存、恢复或重放\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>上下文隔离\u003C\u002Fb>\u003C\u002Fi>：子智能体和工具如何避免相互污染工作状态\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>沙箱与产物\u003C\u002Fb>\u003C\u002Fi>：文件系统、Shell、浏览器及输出结果如何成为上下文管道的一部分\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>人工审批与中断\u003C\u002Fb>\u003C\u002Fi>：生产环境中的智能体在执行高风险或长时间任务时如何保持可控性\u003C\u002Fli>\n\u003C\u002Ful>\n\n### 1. 检索增强生成（RAG）\n\n\u003Cb>综述\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>面向大型语言模型的检索增强生成：综述\u003C\u002Fb>\u003C\u002Fi>, 高云帆等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.10997\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.12-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FTongji-KGLLM\u002FRAG-Survey\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FTongji-KGLLM\u002FRAG-Survey.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>面向定制化大型语言模型的图式检索增强生成综述\u003C\u002Fb>\u003C\u002Fi>, 赵思远等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.13958\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.01-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FDEEP-PolyU\u002FAwesome-GraphRAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FDEEP-PolyU\u002FAwesome-GraphRAG.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>检索增强生成（RAG）及其扩展：关于如何更明智地让LLM使用外部数据的综合综述\u003C\u002Fb>\u003C\u002Fi>, 赵思远等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2409.14924\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.09-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>检索增强生成评估：综述\u003C\u002Fb>\u003C\u002Fi>, 于浩等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.07437\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.07-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FYHPeter\u002FAwesome-RAG-Evaluation\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FYHPeter\u002FAwesome-RAG-Evaluation.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>面向知识密集型NLP任务的检索增强生成\u003C\u002Fb>\u003C\u002Fi>, 刘易斯等, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2005.11401\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2020.05-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fcostadev00\u002FRAG-paper-implementation-from-scratch\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fcostadev00\u002FRAG-paper-implementation-from-scratch.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>面向知识导向的检索增强生成综述\u003C\u002Fb>\u003C\u002Fi>, 程等人, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2503.10677\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.02-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FUSTCAGI\u002FAwesome-Papers-Retrieval-Augmented-Generation\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FUSTCAGI\u002FAwesome-Papers-Retrieval-Augmented-Generation.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>RAG与LLM融合的综述：迈向检索增强型大型语言模型\u003C\u002Fb>\u003C\u002Fi>, 丁等人, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.06211\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.06-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n\u003Cb>朴素RAG\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>超越极限：大型语言模型上下文长度扩展技术综述\u003C\u002Fb>\u003C\u002Fi>, 王新迪等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.02244\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.02-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>机器翻译中的上下文示例选择\u003C\u002Fb>\u003C\u002Fi>, 斯维塔·阿格拉瓦尔等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2212.02437\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>长上下文语言模型时代对RAG的辩护\u003C\u002Fb>\u003C\u002Fi>, 谭宇等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2409.01666\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.09-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>面向知识密集型NLP任务的检索增强生成\u003C\u002Fb>\u003C\u002Fi>, 帕特里克·刘易斯等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2005.11401\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2020.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LightRAG：简单快速的检索增强生成\u003C\u002Fb>\u003C\u002Fi>, 郭子睿等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.05779\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fanonymous.4open.science\u002Fr\u002FLightRAG-2BEE\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fanonymous\u002FLightRAG.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>生成而非检索：大型语言模型是强大的上下文生成器\u003C\u002Fb>\u003C\u002Fi>, 于文浩等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2209.10063\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.09-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fwyu97\u002FGenRead\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fwyu97\u002FGenRead.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>大型语言模型容易被无关上下文分散注意力\u003C\u002Fb>\u003C\u002Fi>, 史弗雷达等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2302.00093\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fgoogle-research-datasets\u002FGSM-IC\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fgoogle-research-datasets\u002FGSM-IC.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>旧式信息检索方法与RAG相遇\u003C\u002Fb>\u003C\u002Fi>, 奥兹·胡利等\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>开放域问答中的密集段落检索\u003C\u002Fb>\u003C\u002Fi>, 弗拉基米尔·卡尔普金等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2004.04906\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2020.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Ffacebookresearch\u002FDPR\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Ffacebookresearch\u002FDPR.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n\u003Cb>高级RAG\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>自适应-RAG：通过问题复杂度学习适应检索增强型大语言模型\u003C\u002Fb>\u003C\u002Fi>, Soyeong Jeong 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2403.14403\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fstarsuzi\u002FAdaptive-RAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fstarsuzi\u002FAdaptive-RAG.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>通过从数万亿个标记中检索来改进语言模型\u003C\u002Fb>\u003C\u002Fi>, Sebastian Borgeaud 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2112.04426\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>FoRAG：面向网络增强型长篇问答的事实性优化检索增强生成\u003C\u002Fb>\u003C\u002Fi>, Tianchi Cai 等人\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>IM-RAG：通过学习内心独白实现多轮检索增强生成\u003C\u002Fb>\u003C\u002Fi>, Diji Yang 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.13021\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>RAGCache：用于检索增强生成的高效知识缓存\u003C\u002Fb>\u003C\u002Fi>, Chao Jin 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.12457\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>纠正型检索增强生成\u003C\u002Fb>\u003C\u002Fi>, Shi-Qi Yan 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.15884\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FHuskyInSalt\u002FCRAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FHuskyInSalt\u002FCRAG.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>RankRAG：在大语言模型中统一上下文排序与检索增强生成\u003C\u002Fb>\u003C\u002Fi>, Yue Yu 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.02485\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Astute RAG：克服大语言模型中的不完美检索增强与知识冲突\u003C\u002Fb>\u003C\u002Fi>, Fei Wang 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.07176\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>学习为检索增强生成过滤上下文\u003C\u002Fb>\u003C\u002Fi>, Zhiruo Wang 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2311.08377\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fzorazrw\u002Ffilco\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fzorazrw\u002Ffilco.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>检索增强型大语言模型中的查询重写\u003C\u002Fb>\u003C\u002Fi>, Xinbei Ma 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2305.14283\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fqijimrc\u002FROBUST\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fqijimrc\u002FROBUST.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>UPRISE：通用提示检索以提升零样本评估\u003C\u002Fb>\u003C\u002Fi>, Daixuan Cheng 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2303.08518\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FMatthewKKai\u002FSMRC2\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FMatthewKKai\u002FSMRC2.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Longllmlingua：通过提示压缩加速和增强大语言模型在长上下文场景中的表现\u003C\u002Fb>\u003C\u002Fi>, Huiqiang Jiang 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.06839\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002FLLMLingua\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmicrosoft\u002FLLMLingua.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>基于条件生成的文档级事件论元抽取\u003C\u002Fb>\u003C\u002Fi>, Sha Li 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2104.05919\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2021.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fraspberryice\u002Fgen-arg\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fraspberryice\u002Fgen-arg.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>多句论元链接\u003C\u002Fb>\u003C\u002Fi>, Seth Ebner 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F1911.03766\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2019.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fnlp.jhu.edu\u002Frams\u002F\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fnlp-jhu\u002FRAMS.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>微调还是检索？比较大语言模型中的知识注入\u003C\u002Fb>\u003C\u002Fi>, Oded Ovadia 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.05934\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>IAG：用于回答推理问题的归纳增强生成框架\u003C\u002Fb>\u003C\u002Fi>, Zhebin Zhang 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2311.18397\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>检索遇见长上下文大语言模型\u003C\u002Fb>\u003C\u002Fi>, Peng Xu 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.03025\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>密集检索 vs. 检索：我们应使用何种检索粒度？\u003C\u002Fb>\u003C\u002Fi>, Tong Chen 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.06648\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fct123098\u002Ffactoid-wiki\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fct123098\u002Ffactoid-wiki.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>利用检索增强探究大语言模型的事实性知识边界\u003C\u002Fb>\u003C\u002Fi>, Ruiyang Ren 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2307.11019\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FRUCAIBox\u002FLLM-Knowledge-Boundary\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FRUCAIBox\u002FLLM-Knowledge-Boundary.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>噪声的力量：重新定义RAG系统的检索\u003C\u002Fb>\u003C\u002Fi>, Florin Cuconasu 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.14887\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fflorin-git\u002FThe-Power-of-Noise\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fflorin-git\u002FThe-Power-of-Noise.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>背诵增强型语言模型\u003C\u002Fb>\u003C\u002Fi>, Zhiqing Sun 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.01296\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FEdward-Sun\u002FRECITE\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FEdward-Sun\u002FRECITE.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>用于零样本槽位填充的鲁棒检索增强生成\u003C\u002Fb>\u003C\u002Fi>, Michael Glass 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2108.13934\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2021.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FIBM\u002Fkgi-slot-filling\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FIBM\u002Fkgi-slot-filling.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>上下文内检索增强型语言模型\u003C\u002Fb>\u003C\u002Fi>, Ori Ram 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2302.00083\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FAI21Labs\u002Fin-context-ralm\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FAI21Labs\u002Fin-context-ralm.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>学习为大语言模型检索上下文示例\u003C\u002Fb>\u003C\u002Fi>, Liang Wang 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2307.07164\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.03-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002FLMOps\u002Ftree\u002Fmain\u002Fllm_retriever\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmicrosoft\u002FLMOps.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>模块化RAG\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>FlashRAG：用于高效检索增强生成研究的模块化工具包\u003C\u002Fb>\u003C\u002Fi>, 金家杰等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.13576\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FRUC-NLPIR\u002FFlashRAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FRUC-NLPIR\u002FFlashRAG.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>多头RAG：利用大语言模型解决多方面问题\u003C\u002Fb>\u003C\u002Fi>, 马切伊·贝斯塔等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2406.05085\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fspcl\u002FMRAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fspcl\u002FMRAG.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>StructRAG：通过推理时混合信息结构化提升大语言模型的知识密集型推理能力\u003C\u002Fb>\u003C\u002Fi>, 李卓群等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.08815\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FLi-Z-Q\u002FStructRAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FLi-Z-Q\u002FStructRAG.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>RAFT：将语言模型适配到领域特定的RAG\u003C\u002Fb>\u003C\u002Fi>, 张天俊等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2403.10131\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FShishirPatil\u002Fgorilla\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FShishirPatil\u002Fgorilla.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>端到端任务导向对话系统的检索-生成对齐\u003C\u002Fb>\u003C\u002Fi>, 沈伟周等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.08877\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fshenwzh3\u002FMK-TOD\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fshenwzh3\u002FMK-TOD.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>UniMS-RAG：面向个性化对话系统的统一多源检索增强生成框架\u003C\u002Fb>\u003C\u002Fi>, 王洪儒等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.13256\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>检索与采样：基于混合检索增强的文档级事件论元抽取\u003C\u002Fb>\u003C\u002Fi>, 任宇兵等。\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>RA-DIT：检索增强的双重指令微调\u003C\u002Fb>\u003C\u002Fi>, 林希维多利亚等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.01352\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Ffacebookresearch\u002FRA-DIT\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Ffacebookresearch\u002FRA-DIT.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>自我知识引导的大语言模型检索增强\u003C\u002Fb>\u003C\u002Fi>, 王一乐等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.05002\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FTHUNLP-MT\u002FSKR\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FTHUNLP-MT\u002FSKR.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>提示引导的非知识密集型任务检索增强\u003C\u002Fb>\u003C\u002Fi>, 郭志成等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2305.17653\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FTHUNLP-MT\u002FPGRA\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FTHUNLP-MT\u002FPGRA.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>REPLUG：检索增强的黑盒语言模型\u003C\u002Fb>\u003C\u002Fi>, 史伟嘉等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2301.12652\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>检索增强型大语言模型的查询重写\u003C\u002Fb>\u003C\u002Fi>, 马新北等， \u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.18653\u002Fv1\u002F2023.emnlp-main.323\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FEMNLP-2023.00-blue\" alt=\"DOI徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fxbmxb\u002FRAG-query-rewriting\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fxbmxb\u002FRAG-query-rewriting.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>自力更生：基于自我记忆的检索增强文本生成\u003C\u002Fb>\u003C\u002Fi>, 成鑫等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2305.02437\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FHannibal046\u002FSelfMemory\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FHannibal046\u002FSelfMemory.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>改进开放域问答中检索增强生成（RAG）模型的领域适应性\u003C\u002Fb>\u003C\u002Fi>, 萨曼·西里瓦德纳等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.02627\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n\u003Cb>基于图的RAG\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>别忘了连接！基于图的重排序提升 RAG\u003C\u002Fb>\u003C\u002Fi>, 董嘉林等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.18414\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>从局部到全局：面向查询摘要的图 RAG 方法\u003C\u002Fb>\u003C\u002Fi>, 达伦·埃奇等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.16130\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>GRAG：图增强生成\u003C\u002Fb>\u003C\u002Fi>, 胡云通等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.16506\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FHuieL\u002FGRAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FHuieL\u002FGRAG.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Iseeq：利用动态元信息检索与知识图谱生成信息查询问题\u003C\u002Fb>\u003C\u002Fi>, 马纳斯·高尔等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2112.07622\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.06-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmanasgaur\u002FAAAI-22\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmanasgaur\u002FAAAI-22.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>G-retriever：用于文本图理解和问答的检索增强生成\u003C\u002Fb>\u003C\u002Fi>, 何晓欣等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.07630\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FXiaoxinHe\u002FG-Retriever\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FXiaoxinHe\u002FG-Retriever.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>面向多文档问答的知识图谱提示\u003C\u002Fb>\u003C\u002Fi>, 王宇等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.08774\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FYuWVandy\u002FKG-LLM-MDQA\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FYuWVandy\u002FKG-LLM-MDQA.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>GNN-RAG：用于大语言模型推理的图神经网络检索\u003C\u002Fb>\u003C\u002Fi>, 科斯塔斯·马夫罗马蒂斯等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.20139\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fcmavro\u002FGNN-RAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fcmavro\u002FGNN-RAG.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LightPROF：知识图谱上大语言模型的轻量级推理框架\u003C\u002Fb>\u003C\u002Fi>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Ftsinghua-fib-lab\u002FACL24-EconAgent\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Ftsinghua-fib-lab\u002FACL24-EconAgent.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>简单即有效：图与大语言模型在基于知识图谱的检索增强生成中的作用\u003C\u002Fb>\u003C\u002Fi>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FGraph-COM\u002FSubgraphRAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FGraph-COM\u002FSubgraphRAG.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>知识图谱引导的检索增强生成\u003C\u002Fb>\u003C\u002Fi>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fnju-websoft\u002FKG2RAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fnju-websoft\u002FKG2RAG.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>MedRAG：通过知识图谱启发式推理增强医疗助手的检索增强生成\u003C\u002Fb>\u003C\u002Fi>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FSNOWTEAM2023\u002FMedRAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FSNOWTEAM2023\u002FMedRAG.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>通过自主知识图谱改造缓解大语言模型幻觉\u003C\u002Fb>\u003C\u002Fi>, KGR 等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2311.13314\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.11-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmansicer\u002FMAIC\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmansicer\u002FMAIC.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>统一框架下基于图的 RAG 深度分析\u003C\u002Fb>\u003C\u002Fi>\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2503.04338\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FJayLZhou\u002FGraphRAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FJayLZhou\u002FGraphRAG.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>RAPTOR：面向树状组织检索的递归抽象处理\u003C\u002Fb>\u003C\u002Fi>, 帕尔斯·萨尔蒂等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.18059\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fparthsarthi03\u002Fraptor\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fparthsarthi03\u002Fraptor.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>TableRAG：使用语言模型进行百万标记表格理解\u003C\u002Fb>\u003C\u002Fi>, 陈思安等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.04739\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fgoogle-research\u002Fgoogle-research\u002Ftree\u002Fmaster\u002Ftable_rag\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fgoogle-research\u002Fgoogle-research.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>KAG：通过知识增强生成提升专业领域中的 LLM\u003C\u002Fb>\u003C\u002Fi>, 梁磊等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2409.13731\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FOpenSPG\u002FKAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FOpenSPG\u002FKAG.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>GFM-RAG：用于检索增强生成的图基础模型\u003C\u002Fb>\u003C\u002Fi>, 罗等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.01113\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FRManLuo\u002Fgfm-rag\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FRManLuo\u002Fgfm-rag.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>HybridRAG：结合向量和图搜索的混合检索系统\u003C\u002Fb>\u003C\u002Fi>, Sarabesh， \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FGitHub-2024.12-white\" alt=\"GitHub Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fsarabesh\u002FHybridRAG\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fsarabesh\u002FHybridRAG.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>代理式 RAG\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>RAG 到记忆：大型语言模型的非参数持续学习\u003C\u002Fb>\u003C\u002Fi>, Bernal Jiménez Gutiérrez 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.14802\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FOSU-NLP-Group\u002FHippoRAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FOSU-NLP-Group\u002FHippoRAG.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>HippoRAG：受神经生物学启发的大型语言模型长期记忆\u003C\u002Fb>\u003C\u002Fi>, Bernal Jiménez Gutiérrez 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.14924\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FOSU-NLP-Group\u002FHippoRAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FOSU-NLP-Group\u002FHippoRAG.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>GraphReader：构建基于图的代理以增强大型语言模型的长上下文能力\u003C\u002Fb>\u003C\u002Fi>, Shilong Li 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2406.14550\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>PlanRAG：一种用于生成式大型语言模型作为决策者的“先规划后检索”增强生成方法\u003C\u002Fb>\u003C\u002Fi>, Myeonghwa Lee 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2406.12430\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmyeon9h\u002FPlanRAG\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmyeon9h\u002FPlanRAG.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Self-RAG：通过自我反思学习检索、生成和批判\u003C\u002Fb>\u003C\u002Fi>, Akari Asai 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.08353\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FAkariAsai\u002Fself-rag\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FAkariAsai\u002Fself-rag.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>DeepRAG：为大型语言模型设计的逐步思考式检索方法\u003C\u002Fb>\u003C\u002Fi>, Xinyan Guan 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.01142\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Paperqa：用于科学研究的检索增强生成式代理\u003C\u002Fb>\u003C\u002Fi>, Jakub Lála 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.07559\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>大型语言模型作为个性化知识驱动对话的源规划器\u003C\u002Fb>\u003C\u002Fi>, Hongru Wang 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2308.06181\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fhrwise-nlp\u002FSAFARI\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fhrwise-nlp\u002FSAFARI.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>PRCA：通过可插拔的奖励驱动上下文适配器，为黑盒大型语言模型适配检索问答任务\u003C\u002Fb>\u003C\u002Fi>, Haoyan Yang 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.18347\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fxbmxb\u002FRAG-query-rewriting\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fxbmxb\u002FRAG-query-rewriting.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>SELF-RAG：通过自我反思学习检索、生成和批判\u003C\u002Fb>\u003C\u002Fi>, Akari Asai 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.11511\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fselfrag.github.io\u002F\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fselfrag\u002Fselfrag.github.io.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>RAT：检索增强思维在长时序生成中激发情境感知推理\u003C\u002Fb>\u003C\u002Fi>, Zihao Wang 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2403.05313\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FCraftJarvis\u002FRAT\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FCraftJarvis\u002FRAT.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>验证链可减少大型语言模型中的幻觉现象\u003C\u002Fb>\u003C\u002Fi>, Shehzaad Dhuliawala 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2309.11495\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.00-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>HM-RAG：层次化多智能体多模态检索增强生成\u003C\u002Fb>\u003C\u002Fi>, Liu 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2504.12330\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Focean-luna\u002FHMRAG\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Focean-luna\u002FHMRAG.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>MultiHop-RAG：面向多跳查询的检索增强生成基准测试\u003C\u002Fb>\u003C\u002Fi>, Tang 和 Yang， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.15391\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.01-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fyixuantt\u002FMultiHop-RAG\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fyixuantt\u002FMultiHop-RAG.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>MMOA-RAG：通过多智能体强化学习改进检索增强生成\u003C\u002Fb>\u003C\u002Fi>, Chen 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2010.10110\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2021.01-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fchenyiqun\u002FMMOA-RAG\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fchenyiqun\u002FMMOA-RAG.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>链中搜索：迈向准确、可信且与时俱进的大型语言模型\u003C\u002Fb>\u003C\u002Fi>, Menick 等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2304.14732\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.04-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>实时与流式 RAG\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>StreamingRAG：实时上下文检索与生成框架\u003C\u002Fb>\u003C\u002Fi>, Sankaradas 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.14101\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.01-red\" alt=\"arXiv 标志\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fvideo-db\u002FStreamRAG\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fvideo-db\u002FStreamRAG.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>面向特定领域且高效的 RAG 的多任务检索器微调\u003C\u002Fb>\u003C\u002Fi>, 作者，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.04652\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.01-red\" alt=\"arXiv 标志\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n\n### 2. 内存系统\n\n#### 运行时内存设计模式\n\n现代内存系统已不再是一个单一的检索存储。生产级智能体越来越多地将以下内容分离：\n\n- **会话\u002F线程状态**，用于当前正在进行的工作\n- **长期语义记忆**，用于存储用户或项目相关的事实\n- **情景记忆**，用于记录轨迹、过往行为和可重用的经验\n- **程序性记忆**，用于存储学习到的工作流程、指令以及稳定的运行偏好\n\n\u003Cb>内存设计参考\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>LangGraph 内存概览\u003C\u002Fb>\u003C\u002Fi>, LangChain，\u003Ca href=\"https:\u002F\u002Fdocs.langchain.com\u002Foss\u002Fjavascript\u002Flanggraph\u002Fmemory\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLangChain-2026-blue\" alt=\"LangChain 标志\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Letta 内存块\u003C\u002Fb>\u003C\u002Fi>, Letta，\u003Ca href=\"https:\u002F\u002Fdocs.letta.com\u002Fguides\u002Fcore-concepts\u002Fmemory\u002Fmemory-blocks\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLetta-2026-blue\" alt=\"Letta 标志\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Claude Code 内存\u003C\u002Fb>\u003C\u002Fi>, Anthropic，\u003Ca href=\"https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fdocs\u002Fclaude-code\u002Fmemory\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAnthropic-2026-blue\" alt=\"Anthropic 标志\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n#### 项目记忆与指令工件\n\n编码智能体已经使项目记忆具体化。在实践中，记忆如今往往以诸如代码库指令文件、限定范围的规则、可重用技能以及长期存在的项目笔记等工件形式存在，而不仅仅局限于向量存储中。\n\n\u003Cb>项目记忆参考\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>介绍 Codex\u003C\u002Fb>\u003C\u002Fi>, OpenAI，\u003Ca href=\"https:\u002F\u002Fopenai.com\u002Findex\u002Fintroducing-codex\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpenAI-2025.05-blue\" alt=\"OpenAI 标志\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Claude Code 内存\u003C\u002Fb>\u003C\u002Fi>, Anthropic，\u003Ca href=\"https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fdocs\u002Fclaude-code\u002Fmemory\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAnthropic-2026-blue\" alt=\"Anthropic 标志\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Claude Code 子代理\u003C\u002Fb>\u003C\u002Fi>, Anthropic，\u003Ca href=\"https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fdocs\u002Fclaude-code\u002Fsub-agents\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAnthropic-2026-blue\" alt=\"Anthropic 标志\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LangChain 深度智能体概览\u003C\u002Fb>\u003C\u002Fi>, LangChain，\u003Ca href=\"https:\u002F\u002Fdocs.langchain.com\u002Foss\u002Fpython\u002Fdeepagents\u002Foverview\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLangChain-2026-blue\" alt=\"LangChain 标志\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>持久化内存架构\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>MemGPT：迈向将 LLM 作为操作系统\u003C\u002Fb>\u003C\u002Fi>, Packer 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.08560\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.10-red\" alt=\"arXiv 标志\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fletta-ai\u002Fletta\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fletta-ai\u002Fletta.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Mem0：构建具有可扩展长期记忆的生产就绪型 AI 智能体\u003C\u002Fb>\u003C\u002Fi>, Taranjeet 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2504.19413\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.04-red\" alt=\"arXiv 标志\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmem0ai\u002Fmem0\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmem0ai\u002Fmem0.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>MemoryLLM：迈向自我更新的大规模语言模型\u003C\u002Fb>\u003C\u002Fi>, Wang 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.04624\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.02-red\" alt=\"arXiv 标志\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fwangyu-ustc\u002FMemoryLLM\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fwangyu-ustc\u002FMemoryLLM.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Infinite-LLM：基于 DistAttention 和分布式 KVCache 的高效长上下文 LLM 服务\u003C\u002Fb>\u003C\u002Fi>, 匿名作者，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.02669\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.01-red\" alt=\"arXiv 标志\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>内存增强型生成对抗变压器\u003C\u002Fb>\u003C\u002Fi>, 匿名作者，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.19218\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.02-red\" alt=\"arXiv 标志\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>内存交换标准\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>PAM（便携式 AI 内存）：一种用于 AI 用户记忆的开放交换格式\u003C\u002Fb>\u003C\u002Fi>, Daniel Gines，\u003Ca href=\"https:\u002F\u002Fportable-ai-memory.org\u002Fspec\u002Fv1.0\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FSpec-v1.0-blue\" alt=\"规范标志\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fportable-ai-memory\u002Fpython-sdk\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fportable-ai-memory\u002Fpython-sdk.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>内存增强型神经网络\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>内存增强型神经网络综述：从认知洞察到 AI 应用\u003C\u002Fb>\u003C\u002Fi>, Khosla 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.06141\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.12-red\" alt=\"arXiv 标志\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>一种具备短期、情景及语义记忆系统的机器\u003C\u002Fb>\u003C\u002Fi>, Kim 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2212.02098\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.12-red\" alt=\"arXiv 标志\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fhumemai\u002Fagent-room-env-v1\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fhumemai\u002Fagent-room-env-v1.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>从人类记忆到 AI 内存：LLM 时代记忆机制综述\u003C\u002Fb>\u003C\u002Fi>, Wu 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2504.15965\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.04-red\" alt=\"arXiv 标志\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>情景记忆与上下文持久性\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>记忆在大语言模型中的作用：持久化上下文以实现更智能的对话\u003C\u002Fb>\u003C\u002Fi>, Porcu, \u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.18535\u002Fijsrm\u002Fv12i11.ec04\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FIJSRM-2024.11-blue\" alt=\"IJSRM徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>AI代理中的情景记忆存在风险，应加以研究和缓解\u003C\u002Fb>\u003C\u002Fi>, Christiano等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.11739\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.10-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Larimar：具有情景记忆控制的大语言模型\u003C\u002Fb>\u003C\u002Fi>, Goyal等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2403.11901\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICML-2024.03-blue\" alt=\"ICML徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>EM-LLM：类人情景记忆用于无限上下文的大语言模型\u003C\u002Fb>\u003C\u002Fi>, 匿名等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.09450\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2024.07-blue\" alt=\"ICLR徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fem-llm\u002FEM-LLM-model\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fem-llm\u002FEM-LLM-model.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>具有可控工作记忆的大语言模型\u003C\u002Fb>\u003C\u002Fi>, Goyal等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2211.05110\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.11-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>增强大语言模型代理的工作记忆\u003C\u002Fb>\u003C\u002Fi>, 匿名等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.17259\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.12-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>持续学习与记忆巩固\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>基于预测误差驱动的记忆巩固用于持续学习\u003C\u002Fb>\u003C\u002Fi>, 匿名等，\u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNeurIPS-2020.11-blue\" alt=\"NeurIPS徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>通过探索海森矩阵的特征值克服持续学习中的灾难性遗忘\u003C\u002Fb>\u003C\u002Fi>, 匿名等，\u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNeurIPS-2023-blue\" alt=\"NeurIPS徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>脉冲网络中利用忆阻器实现持续学习的概率型元可塑性\u003C\u002Fb>\u003C\u002Fi>, 匿名等，\u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.11-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>对话记忆\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>MemoChat：调优大语言模型以使用备忘录进行一致的长程开放域对话\u003C\u002Fb>\u003C\u002Fi>, 匿名等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2308.08239\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.08-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>思考即记忆：回忆与后思使大语言模型具备长期记忆\u003C\u002Fb>\u003C\u002Fi>, 匿名等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2311.08719\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.11-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>生成式代理：人类行为的交互式模拟物\u003C\u002Fb>\u003C\u002Fi>, Park等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2304.03442\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.04-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>大语言模型的自控记忆框架\u003C\u002Fb>\u003C\u002Fi>, 匿名等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2304.13343\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.04-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>个性化与记忆\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>通过参数化用户记忆注入实现个性化大语言模型响应生成\u003C\u002Fb>\u003C\u002Fi>, 匿名等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.03565\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>灵魂驱动的交互设计：关于AI代理声明式人格规范的立场论文\u003C\u002Fb>\u003C\u002Fi>, Lee，\u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.5281\u002Fzenodo.18678616\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FZenodo-2026.02-blue\" alt=\"Zenodo徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Soul Spec——AI代理人格包的开放规范\u003C\u002Fb>\u003C\u002Fi>, ClawSouls，\u003Ca href=\"https:\u002F\u002Fclawsouls.ai\u002Fspec\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FSpec-v0.4-blue\" alt=\"Spec徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fclawsouls\u002Fsoul-spec-mcp\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fclawsouls\u002Fsoul-spec-mcp.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>安全与对齐及记忆\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>宪章式AI：来自AI反馈的无害性\u003C\u002Fb>\u003C\u002Fi>, Bai等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2212.08073\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.12-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>通过有针对性的人类判断（Sparrow）改进对话代理的对齐\u003C\u002Fb>\u003C\u002Fi>, Glaese等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2209.14375\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.09-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>工具集成与记忆\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>WebGPT：浏览器辅助问答结合人类反馈\u003C\u002Fb>\u003C\u002Fi>, Nakano等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2112.09332\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2021.12-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>ToolLLM：助力大语言模型掌握16000多种真实世界API\u003C\u002Fb>\u003C\u002Fi>, Qin等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2307.16789\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.07-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>学习与反思\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>语言模型是少样本学习者（GPT-3）\u003C\u002Fb>\u003C\u002Fi>, Brown等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2005.14165\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2020.05-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Reflexion：具备言语强化学习能力的语言代理\u003C\u002Fb>\u003C\u002Fi>, Shinn等，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2303.11366\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNeurIPS-2023.03-blue\" alt=\"NeurIPS徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fnoahshinn\u002Freflexion\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fnoahshinn\u002Freflexion.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n\n### 3. 代理通信\n\n\u003Cb>调查\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>人工智能代理协议综述\u003C\u002Fb>\u003C\u002Fi>, 杨英轩等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2504.16736\" target=\"_blank\">\u003Cimg src=\"https0:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.06-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fzoe-yyx\u002FAwesome-AIAgent-Protocol\" target=\"_blank\">\n        \u003Cimg src=\"https0:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fzoe-yyx\u002FAwesome-AIAgent-Protocol.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>具有通信功能的多智能体深度强化学习综述\u003C\u002Fb>\u003C\u002Fi>, 朱昌熙等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2203.08975\" target=\"_blank\">\u003Cimg src=\"https0:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.10-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>超越自我对话：基于LLM的多智能体系统以通信为中心的综述\u003C\u002Fb>\u003C\u002Fi>, 颜冰宇等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.14321\" target=\"_blank\">\u003Cimg src=\"https0:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.06-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>基于大型语言模型的多智能体：进展与挑战综述\u003C\u002Fb>\u003C\u002Fi>, 郭泰成等， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.01680\" target=\"_blank\">\u003Cimg src=\"https0:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Ftaichengguo\u002FLLM_MultiAgents_Survey_Papers\" target=\"_blank\">\n        \u003Cimg src=\"https0:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Ftaichengguo\u002FLLM_MultiAgents_Survey_Papers.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n#### 开放式代理协议与互操作性\n\n开放式协议已成为代理工程的重要组成部分。在实践中，现代代理系统越来越倾向于将以下方面分离：\n\n- **代理与工具之间的协议**，如MCP\n- **代理与代理之间的协议**，如A2A和ACP风格的远程调用\n- **代理与用户界面之间的协议**，如AG-UI\n- **可移植的代理定义**，如AgentSchema\n\n\u003Cb>官方协议与互操作性参考资料\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>模型上下文协议规范\u003C\u002Fb>\u003C\u002Fi>, MCP工作组， \u003Ca href=\"https:\u002F\u002Fmodelcontextprotocol.io\u002Fspecification\u002F2025-06-18\" target=\"_blank\">\u003Cimg src=\"https0:\u002F\u002Fimg.shields.io\u002Fbadge\u002FSpec-2025.06-blue\" alt=\"规范徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>模型上下文协议架构\u003C\u002Fb>\u003C\u002Fi>, MCP工作组， \u003Ca href=\"https:\u002F\u002Fmodelcontextprotocol.io\u002Fdocs\u002Flearn\u002Farchitecture\" target=\"_blank\">\u003Cimg src=\"https0:\u002F\u002Fimg.shields.io\u002Fbadge\u002FDocs-2026-blue\" alt=\"文档徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Agent2Agent协议（A2A）\u003C\u002Fb>\u003C\u002Fi>, Google， \u003Ca href=\"https:\u002F\u002Fa2a-protocol.org\u002Flatest\u002F\" target=\"_blank\">\u003Cimg src=\"https0:\u002F\u002Fimg.shields.io\u002Fbadge\u002FProtocol-2026-blue\" alt=\"协议徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>AG-UI文档\u003C\u002Fb>\u003C\u002Fi>, CopilotKit团队， \u003Ca href=\"https:\u002F\u002Fdocs.ag-ui.com\u002F\" target=\"_blank\">\u003Cimg src=\"https0:\u002F\u002Fimg.shields.io\u002Fbadge\u002FProtocol-2026-blue\" alt=\"协议徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>ACP Connect\u003C\u002Fb>\u003C\u002Fi>, AGNTCY， \u003Ca href=\"https:\u002F\u002Fdocs.agntcy.org\u002Fsyntactic\u002Fconnect\u002F\" target=\"_blank\">\u003Cimg src=\"https0:\u002F\u002Fimg.shields.io\u002Fbadge\u002FProtocol-2026-blue\" alt=\"协议徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>AgentSchema\u003C\u002Fb>\u003C\u002Fi>, 微软， \u003Ca href=\"https:\u002F\u002Fmicrosoft.github.io\u002FAgentSchema\u002F\" target=\"_blank\">\u003Cimg src=\"https0:\u002F\u002Fimg.shields.io\u002Fbadge\u002FSchema-2026-blue\" alt=\"模式徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>代理互操作协议\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>代理互操作协议综述：模型上下文协议（MCP）、代理通信协议（ACP）和代理间协议（A2A）\u003C\u002Fb>\u003C\u002Fi>, 张等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2505.02279\" target=\"_blank\">\u003Cimg src=\"https0:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.05-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>通过身份感知学习实现富有表现力的多智能体通信\u003C\u002Fb>\u003C\u002Fi>, 杜等人， \u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.1609\u002Faaai.v38i16.29683\" target=\"_blank\">\u003Cimg src=\"https0:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAAAI-2024.03-blue\" alt=\"AAAI徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>面向多智能体强化学习的上下文感知通信（CACOM）\u003C\u002Fb>\u003C\u002Fi>, 李等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.15600\" target=\"_blank\">\u003Cimg src=\"https0:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.12-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FLXXXXR\u002FCACOM\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https0:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FLXXXXR\u002FCACOM.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>代理互操作协议综述：模型上下文协议（MCP）、代理通信协议（ACP）、代理间协议（A2A）和代理网络协议（ANP）\u003C\u002Fb>\u003C\u002Fi>, Abul Ehtesham等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2505.02279\" target=\"_blank\">\u003Cimg src=\"https0:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.05-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>代理能力协商与绑定协议（ACNBP）\u003C\u002Fb>\u003C\u002Fi>, 肯·黄等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2506.13590\" target=\"_blank\">\u003Cimg src=\"https0:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.06-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>用于大型语言模型网络的可扩展通信协议\u003C\u002Fb>\u003C\u002Fi>, Samuele Marro等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.11905\" target=\"_blank\">\u003Cimg src=\"https0:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.10-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fagora-protocol\u002Fpaper-demo\" target=\"_blank\">\n        \u003Cimg src=\"https0:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fagora-protocol\u002Fpaper-demo.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>模型上下文协议（MCP）\u003C\u002Fb>\u003C\u002Fi>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmodelcontextprotocol\u002Fmodelcontextprotocol\" target=\"_blank\">\n        \u003Cimg src=\"https0:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmodelcontextprotocol\u002Fmodelcontextprotocol.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Agent2Agent（A2A）协议\u003C\u002Fb>\u003C\u002Fi>\n    \u003Ca href=\"https:\u002F\u002Fgoogle.github.io\u002FA2A\u002F\" target=\"_blank\">\n        \u003Cimg src=\"https0:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fgoogle\u002FA2A.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>代理网络协议（ANP）\u003C\u002Fb>\u003C\u002Fi>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fagent-network-protocol\u002FAgentNetworkProtocol\" target=\"_blank\">\n        \u003Cimg src=\"https0:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fagent-network-protocol\u002FAgentNetworkProtocol.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n\u003Cb>结构化通信框架\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>面向多智能体强化学习的结构化通信学习\u003C\u002Fb>\u003C\u002Fi>, Wang 等人，\u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAAMAS-2023.05-blue\" alt=\"AAMAS徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fbellmanequation\u002FLSC\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fbellmanequation\u002FLSC.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>AC2C：用于多智能体强化学习的自适应两跳通信\u003C\u002Fb>\u003C\u002Fi>, Wang 等人，\u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAAMAS-2023.05-blue\" alt=\"AAMAS徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>面向智能体间通信的任务无关对比预训练\u003C\u002Fb>\u003C\u002Fi>, Sun 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.02174\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAAMAS-2025.05-blue\" alt=\"AAMAS徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>AC2C：用于多智能体强化学习的自适应两跳通信\u003C\u002Fb>\u003C\u002Fi>, Xuefeng Wang 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2302.12515\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.12-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>CAMEL：用于大型语言模型社会“心智”探索的通信智能体\u003C\u002Fb>\u003C\u002Fi>, Guohao Li 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2303.17760\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.11-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fcamel-ai\u002Fcamel\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fcamel-ai\u002Fcamel.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>面向多智能体强化学习的上下文感知通信（CACOM）\u003C\u002Fb>\u003C\u002Fi>, Xinran Li 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.15600\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.07-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FLXXXXR\u002FCACOM\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FLXXXXR\u002FCACOM.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>面向智能体间通信的任务无关对比预训练\u003C\u002Fb>\u003C\u002Fi>, Peihong Yu 等人\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>基于身份感知学习的富有表现力的多智能体通信\u003C\u002Fb>\u003C\u002Fi>, Wei Du 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.07872\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.03-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>MAGIS：基于LLM的多智能体框架，用于解决GitHub问题\u003C\u002Fb>\u003C\u002Fi>, Wei Tao 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2403.17927\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.06-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>AutoAgents：自动智能体生成框架\u003C\u002Fb>\u003C\u002Fi>, Guangyao Chen 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2309.17288\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FLink-AGI\u002FAutoAgents\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FLink-AGI\u002FAutoAgents.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>MDTeamGPT：一种自我演进的基于LLM的多智能体框架，用于多学科团队医疗会诊\u003C\u002Fb>\u003C\u002Fi>, Kai Chen 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2503.13856\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.03-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fkaichennj.github.io\u002FMDTeamGPT\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fkaichennj\u002FMDTeamGPT.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>AutoGen：通过多智能体对话实现下一代LLM应用\u003C\u002Fb>\u003C\u002Fi>, Wu 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2308.08155\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.08-red\" alt=\"arXiv徽章\">\u003C\u002Fa> \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002Fautogen\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmicrosoft\u002Fautogen.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\n\u003Cb>LLM增强的智能体通信\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>ProAgent：利用大型語言模型構建主動協作型智能體\u003C\u002Fb>\u003C\u002Fi>, Ceyao Zhang 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2308.11339\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.01-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fpku-proagent.github.io\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fpku-proagent\u002Fproagent.svg?style=social\" alt=\"GitHub星標\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>通過多智能體辯論提升語言模型的事實性和推理能力\u003C\u002Fb>\u003C\u002Fi>, Yilun Du 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2305.14325\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.05-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fcomposable-models.github.io\u002Fllm_debate\u002F\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fcomposable-models\u002Fllm_debate.svg?style=social\" alt=\"GitHub星標\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>ChatDev：用於軟體開發的溝通型智能體\u003C\u002Fb>\u003C\u002Fi>, Chen Qian 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2307.07924\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.06-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FOpenBMB\u002FChatDev\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FOpenBMB\u002FChatDev.svg?style=social\" alt=\"GitHub星標\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>基於去中心化隊友建模的多智能體激勵性溝通\u003C\u002Fb>\u003C\u002Fi>, Nian Li 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.10436\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Ftsinghua-fib-lab\u002FACL24-EconAgent\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Ftsinghua-fib-lab\u002FACL24-EconAgent.svg?style=social\" alt=\"GitHub星標\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>AgentCoord：基於LLM的多智能體協作協調策略可視化探索\u003C\u002Fb>\u003C\u002Fi>, Bo Pan 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.11943\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FAgentCoord\u002FAgentCoord\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FAgentCoord\u002FAgentCoord.svg?style=social\" alt=\"GitHub星標\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>基於LLM智能體網絡模擬意見動態\u003C\u002Fb>\u003C\u002Fi>, Yun-Shiuan Chuang 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2311.09618\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fyunshiuan\u002Fllm-agent-opinion-dynamics\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fyunshiuan\u002Fllm-agent-opinion-dynamics.svg?style=social\" alt=\"GitHub星標\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>MetaGPT：面向多智能體協作框架的元編程\u003C\u002Fb>\u003C\u002Fi>, Sirui Hong 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2308.00352\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.11-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fgeekan\u002FMetaGPT\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fgeekan\u002FMetaGPT.svg?style=social\" alt=\"GitHub星標\">\n    \u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>智能體鏈：大型語言模型在長上下文任務中的協作\u003C\u002Fb>\u003C\u002Fi>, Yusen Zhang 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2406.02818\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.06-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>基於去中心化隊友建模的多智能體激勵性溝通\u003C\u002Fb>\u003C\u002Fi>, Lei Yuan 等人。\u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.1609\u002Faaai.v36i9.21179\" target=\"_blank\">\n        \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAAAI-2022.06-blue\" alt=\"DOI徽章\">\n\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>ProAgent：利用大型語言模型構建主動協作型智能體\u003C\u002Fb>\u003C\u002Fi>, Zhang 等人，\u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.1609\u002Faaai.v38i16.29710\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAAAI-2024.03-blue\" alt=\"AAAI徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FPKU-Alignment\u002FProAgent\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FPKU-Alignment\u002FProAgent.svg?style=social\" alt=\"GitHub星標\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>模型上下文協議（MCP）\u003C\u002Fb>\u003C\u002Fi>, Anthropic，\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmodelcontextprotocol\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FGitHub-2024-white\" alt=\"GitHub徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>CoMAS：通過交互獎勵實現多智能體系統的共同演化\u003C\u002Fb>\u003C\u002Fi>, Xue 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2510.08529\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.10-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>分布式多智能體系統的阿喀琉斯之踵\u003C\u002Fb>\u003C\u002Fi>, Zhang 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2504.07461\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.04-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n### 4. 工具使用与函数调用\n\n#### 托管代理工具与计算机使用\n\n工具使用的前沿已从静态的函数模式，转向**托管工具运行时**、**远程服务器**以及**计算机使用界面**。在代理时代，工具越来越多地通过平台管理的执行流程、审批流和UI感知的控制回路来连接，而非单次的JSON调用。\n\n\u003Cb>官方工具与计算机使用参考资料\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>OpenAI 工具指南\u003C\u002Fb>\u003C\u002Fi>, OpenAI, \u003Ca href=\"https:\u002F\u002Fdevelopers.openai.com\u002Fapi\u002Fdocs\u002Fguides\u002Ftools\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpenAI-2026-blue\" alt=\"OpenAI Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>介绍 Codex\u003C\u002Fb>\u003C\u002Fi>, OpenAI, \u003Ca href=\"https:\u002F\u002Fopenai.com\u002Findex\u002Fintroducing-codex\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpenAI-2025.05-blue\" alt=\"OpenAI Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Claude 3.5 的计算机使用功能\u003C\u002Fb>\u003C\u002Fi>, Anthropic, \u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fnews\u002F3-5-models-and-computer-use\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAnthropic-2024.10-blue\" alt=\"Anthropic Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Google Vertex AI 代理引擎\u003C\u002Fb>\u003C\u002Fi>, Google, \u003Ca href=\"https:\u002F\u002Fcloud.google.com\u002Fvertex-ai\u002Fgenerative-ai\u002Fdocs\u002Fagent-engine\u002Foverview\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FGoogle-2026-blue\" alt=\"Google Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>OSWorld\u003C\u002Fb>\u003C\u002Fi>, Xie 等人, \u003Ca href=\"https:\u002F\u002Fos-world.github.io\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FBenchmark-2026-orange\" alt=\"Benchmark Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Lumen\u003C\u002Fb>\u003C\u002Fi> — 以视觉为核心的浏览器代理，具备基于CDP的自愈性确定性回放能力。采用截图 → 模型 → 动作循环，并支持多提供商（Anthropic、Google）。 \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fomxyz\u002Flumen\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fomxyz\u002Flumen.svg?style=social\" alt=\"GitHub 星标\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>基础工具学习\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Toolformer：语言模型可自我学习如何使用工具\u003C\u002Fb>\u003C\u002Fi>, Schick 等人, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2302.04761\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNeurIPS-2023.09-blue\" alt=\"NeurIPS Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fxrsrke\u002Ftoolformer\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fxrsrke\u002Ftoolformer.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>ReAct：在语言模型中协同推理与行动\u003C\u002Fb>\u003C\u002Fi>, Yao 等人, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.03629\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2022.10-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fysymyth\u002FReAct\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fysymyth\u002FReAct.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>增强型语言模型：综述\u003C\u002Fb>\u003C\u002Fi>, Qin 等人, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2302.07842\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.02-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>大型语言模型中的工具学习：综述\u003C\u002Fb>\u003C\u002Fi>, Qu 等人, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.17935\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.05-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fquchangle1\u002FLLM-Tool-Survey\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fquchangle1\u002FLLM-Tool-Survey.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\u003Cb>高级函数调用系统\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Granite 函数调用模型：通过细粒度任务的多任务学习引入函数调用能力\u003C\u002Fb>\u003C\u002Fi>, Smith 等人, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.00121\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.06-red\" alt=\"arXiv Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>HuggingGPT：利用 ChatGPT 及其 Hugging Face 伙伴解决 AI 任务\u003C\u002Fb>\u003C\u002Fi>, Shen 等人, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2303.17580\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNeurIPS-2023.09-blue\" alt=\"NeurIPS Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fquchangle1\u002Fmicrosoft\u002FJARVIS\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fmicrosoft\u002FJARVIS.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>提升 LLM 中的函数调用能力：提示格式、数据集成与多语言翻译策略\u003C\u002Fb>\u003C\u002Fi>, Chen 等人, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.01130\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNAACL-2025.04-blue\" alt=\"NAACL Badge\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>用于复杂网络任务的真实世界 WebAgent\u003C\u002Fb>\u003C\u002Fi>, Zhai 等人, \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fhtml\u002F2307.12856v4\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.03-red\" alt=\"arXiv Badge\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\u003Cb>多智能体函数调用\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>ToolACE：赢得 LLM 函数调用竞赛\u003C\u002Fb>\u003C\u002Fi>, Zhang 等人, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpenReview-2025.03-orange\" alt=\"OpenReview Badge\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>伯克利函数排行榜（BFCL）：评估函数调用能力\u003C\u002Fb>\u003C\u002Fi>, 各方, \u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FBenchmark-2024-orange\" alt=\"Benchmark Badge\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FShishirPatil\u002Fgorilla\u002Ftree\u002Fmain\u002Fberkeley-function-call-leaderboard\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FShishirPatil\u002Fgorilla.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n---\n\n## 📊 上下文驱动系统的评估范式\n\n### 上下文质量评估\n\n\u003Cb>基础长上下文基准\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>RULER：你的长上下文语言模型的真实上下文大小是多少？\u003C\u002Fb>\u003C\u002Fi>, Cheng-Ping Hsieh 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.06654\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FCOLM-2024.07-blue\" alt=\"COLM徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FRULER\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FNVIDIA\u002FRULER.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LongBench：一个双语、多任务的长上下文理解基准\u003C\u002Fb>\u003C\u002Fi>, Bai 等人，\u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FACL-2024.08-blue\" alt=\"ACL徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>∞BENCH：将长上下文评估扩展到10万标记之外\u003C\u002Fb>\u003C\u002Fi>, Zhang 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.13718\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FACL-2024.08-blue\" alt=\"ACL徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FTHUDM\u002FLongBench\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FTHUDM\u002FLongBench.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>VL-ICL基准：多模态上下文学习中的细节魔鬼\u003C\u002Fb>\u003C\u002Fi>, Zong 等人，\u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FICLR-2025.01-blue\" alt=\"ICLR徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fys-zong\u002FVL-ICL\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fys-zong\u002FVL-ICL.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\u003Cb>多模态与专项评估\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>多模态针在 haystack 中：多模态大型语言模型的长上下文能力基准测试\u003C\u002Fb>\u003C\u002Fi>, Wang 等人，\u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNAACL-2025.04-blue\" alt=\"NAACL徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FWang-ML-Lab\u002Fmultimodal-needle-in-a-haystack\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FWang-ML-Lab\u002Fmultimodal-needle-in-a-haystack.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>情境化主题连贯性（CTC）指标\u003C\u002Fb>\u003C\u002Fi>, Rahimi 等人，\u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FACL-2024.03-blue\" alt=\"ACL徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FhamedR96\u002FCTC\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FhamedR96\u002FCTC.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>BBScore：一种基于布朗桥的文本连贯性评估指标\u003C\u002Fb>\u003C\u002Fi>, Sheng 等人，\u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.1609\u002Faaai.v38i13.29414\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAAAI-2024.03-blue\" alt=\"AAAI徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fzcsheng95\u002FBBScore\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fzcsheng95\u002FBBScore.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n\u003Cb>RAG与生成评估\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>检索增强生成评估：综述\u003C\u002Fb>\u003C\u002Fi>, Li 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.07437\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.05-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Ragas：检索增强生成的自动化评估\u003C\u002Fb>\u003C\u002Fi>, Espinosa-Anke 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2309.15217\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.09-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>临床微生物学中生成式AI聊天机器人的人工评估协议\u003C\u002Fb>\u003C\u002Fi>, Griego-Herrera 等人，\u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.1371\u002Fjournal.pone.0300487\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FPLOS-2024.03-blue\" alt=\"PLOS徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n### 上下文工程基准测试\n\n\u003Cb>合成与现实评估\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>针在 haystack 中（NIAH）及合成基准\u003C\u002Fb>\u003C\u002Fi>, 研究领域 2023–2024年，\u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FBenchmark-2024-orange\" alt=\"Benchmark徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fgkamradt\u002FLLMTest_NeedleInAHaystack\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fgkamradt\u002FLLMTest_NeedleInAHaystack.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>ZeroSCROLLS：真实的自然语言任务\u003C\u002Fb>\u003C\u002Fi>, 基准 2023–2024年，\u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FBenchmark-2024-orange\" alt=\"Benchmark徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Ftau-nlp\u002Fzero_scrolls\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Ftau-nlp\u002Fzero_scrolls.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>InfiniteBench：10万+标记的评估\u003C\u002Fb>\u003C\u002Fi>, 基准 2024年，\u003Ca href=\"#\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FBenchmark-2024-orange\" alt=\"Benchmark徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FOpenBMB\u002FInfiniteBench\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FOpenBMB\u002FInfiniteBench.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Agent-Pro：通过基于提案的编程学习进化编码代理\u003C\u002Fb>\u003C\u002Fi>, Zhang 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.17574\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.05-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>GenoTEX：用于自动基因表达数据分析的LLM代理基准\u003C\u002Fb>\u003C\u002Fi>, Liu 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2406.15341\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FMLCB-2025.06-blue\" alt=\"MLCB徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FLiu-Hy\u002FGenoTEX\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FLiu-Hy\u002FGenoTEX.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n### 代理可观测性与遥测\n\n长时间运行的代理系统需要的不仅仅是离线基准测试分数。它们还需要在跟踪级别上，对计划、工具调用、内存读写、审批、重试以及失败模式有清晰的可见性。可观测性正日益成为生产环境中上下文工程的验证层。\n\n\u003Cb>可观测性和遥测相关参考资料\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>LangSmith 可观测性快速入门\u003C\u002Fb>\u003C\u002Fi>, LangChain, \u003Ca href=\"https:\u002F\u002Fdocs.langchain.com\u002Flangsmith\u002Fobservability-quickstart\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLangChain-2026-blue\" alt=\"LangChain 标志\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>生成式 AI 的 OpenTelemetry 语义规范\u003C\u002Fb>\u003C\u002Fi>, OpenTelemetry, \u003Ca href=\"https:\u002F\u002Fopentelemetry.io\u002Fdocs\u002Fspecs\u002Fsemconv\u002Fgen-ai\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpenTelemetry-2026-blue\" alt=\"OpenTelemetry 标志\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Google ADK 评估与可观测性\u003C\u002Fb>\u003C\u002Fi>, Google, \u003Ca href=\"https:\u002F\u002Fgoogle.github.io\u002Fadk-docs\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FGoogle-2026-blue\" alt=\"Google 标志\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>OpenAI 代理与工具\u003C\u002Fb>\u003C\u002Fi>, OpenAI, \u003Ca href=\"https:\u002F\u002Fplatform.openai.com\u002Fdocs\u002Fguides\u002Fagents\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpenAI-2026-blue\" alt=\"OpenAI 标志\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\n---\n\n## 🚀 应用与系统\n\n### 复杂研究系统\n\n\u003Cb>假设生成与数据驱动发现\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>基于大型语言模型的假设生成\u003C\u002Fb>\u003C\u002Fi>, 刘等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.04326\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.04-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FChicagoHAI\u002Fhypothesis-generation\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FChicagoHAI\u002Fhypothesis-generation.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>GFlowNets用于人工智能驱动的科学发现\u003C\u002Fb>\u003C\u002Fi>, 贾因等人， \u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.1039\u002FD3DD00002H\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FDigital_Discovery-2023.06-blue\" alt=\"Digital Discovery徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>文献与数据结合：一种协同的假设生成方法\u003C\u002Fb>\u003C\u002Fi>, 刘等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.17309\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.10-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FChicagoHAI\u002Fhypothesis-generation\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FChicagoHAI\u002Fhypothesis-generation.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>机器学习在生物医学领域的假设生成\u003C\u002Fb>\u003C\u002Fi>, FieldSHIFT团队， \u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.1039\u002FD3DD00185G\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FDigital_Discovery-2024.02-blue\" alt=\"Digital Discovery徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>自动化科学发现\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>AI科学家：迈向完全自动化的开放式科学发现\u003C\u002Fb>\u003C\u002Fi>, 陆等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2408.06292\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.08-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FSakanaAI\u002FAI-Scientist\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FSakanaAI\u002FAI-Scientist.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>利用AI自动化心理学假设生成\u003C\u002Fb>\u003C\u002Fi>, 约翰逊等人， \u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.1038\u002Fs41599-024-03407-5\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNature-2024.07-blue\" alt=\"Nature徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>大型语言模型能否取代人类进行系统性综述？\u003C\u002Fb>\u003C\u002Fi>, 克莱莎等人， \u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.1002\u002Fjrsm.1715\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FResearch_Synthesis-2024-blue\" alt=\"Research Synthesis徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>无需人类示范即可解决奥林匹克几何问题\u003C\u002Fb>\u003C\u002Fi>, 郑等人， \u003Ca href=\"https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41586-023-06747-5\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNature-2024.01-blue\" alt=\"Nature徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>GenoMAS：基于代码驱动基因表达分析的多智能体科学发现框架\u003C\u002Fb>\u003C\u002Fi>, 刘等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.21035\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.07-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FLiu-Hy\u002FGenoMAS\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FLiu-Hy\u002FGenoMAS.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>aiXiv：由AI科学家生成的下一代开放获取科学发现生态系统\u003C\u002Fb>\u003C\u002Fi>, 张等人， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2508.15126\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2025.08-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Faixiv-org\u002FaiXiv\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Faixiv-org\u002FaiXiv.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>AI赋能科学的整合与未来方向\u003C\u002Fb>\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>AI赋能科学2025：AI创新与科学发现的融合\u003C\u002Fb>\u003C\u002Fi>, 芬克等人， \u003Ca href=\"https:\u002F\u002Fdoi.org\u002F10.1038\u002Fd41573-025-00161-3\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FNature-2025.05-blue\" alt=\"Nature徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>迈向生成式AI驱动的科学发现：进展、机遇与挑战\u003C\u002Fb>\u003C\u002Fi>, 匿名作者， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.11427\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2023.12-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>深度研究应用\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>利用AI加速科学发现\u003C\u002Fb>\u003C\u002Fi>, MIT新闻， \u003Ca href=\"https:\u002F\u002Fnews.mit.edu\u002F2025\u002Ffuturehouse-accelerates-scientific-discovery-with-ai-0630\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FMIT-2025.06-blue\" alt=\"MIT徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>借助AI联合科学家加速科学突破\u003C\u002Fb>\u003C\u002Fi>, Google研究， \u003Ca href=\"https:\u002F\u002Fresearch.google\u002Fblog\u002Faccelerating-scientific-breakthroughs-with-an-ai-co-scientist\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FGoogle-2024-blue\" alt=\"Google徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>连接AI与科学：来自大规模AI4Science文献分析的启示\u003C\u002Fb>\u003C\u002Fi>, 各方， \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.09628\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2024.12-red\" alt=\"arXiv徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fcharles-pyj\u002FBridging-AI-and-Science\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fcharles-pyj\u002FBridging-AI-and-Science.svg?style=social\" alt=\"GitHub星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>AI赋能科学发现\u003C\u002Fb>\u003C\u002Fi>, 世界经济论坛， \u003Ca href=\"https:\u002F\u002Fwww.weforum.org\u002Fpublications\u002Ftop-10-emerging-technologies-2024\u002Fin-full\u002F1-ai-for-scientific-discovery\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FWEF-2024-blue\" alt=\"WEF徽章\">\u003C\u002Fa>\n    \u003C\u002Fli>\n\u003C\u002Ful>\n\n### 生产系统\n\n\u003Cb>上下文工程作为核心学科\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>从提示词设计到系统设计：上下文工程作为人工智能驱动交付的核心学科\u003C\u002Fb>\u003C\u002Fi>, Forte Group 团队, \u003Ca href=\"https:\u002F\u002Ffortegrp.com\u002Finsights\u002Fcontext-engineering-as-a-core-discipline-for-ai-driven-delivery\" target=\"_blank\">\u003Cimg src=\"https600;:\u002F\u002Fimg.shields.io\u002Fbadge\u002FForte-2025.07-blue\" alt=\"Forte 徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>上下文工程：企业级 AI 运营框架\u003C\u002Fb>\u003C\u002Fi>, Shelly Palmer, \u003Ca href=\"https:\u002F\u002Fshellypalmer.com\u002F2025\u002F06\u002Fcontext-engineering-a-framework-for-enterprise-ai-operations\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FShellyPalmer-2025.06-blue\" alt=\"ShellyPalmer 徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>MCP 如何在高吞吐量场景下处理上下文管理\u003C\u002Fb>\u003C\u002Fi>, Portkey.ai 团队, \u003Ca href=\"https:\u002F\u002Fportkey.ai\u002Fblog\u002Fmodel-context-protocol-context-management-in-high-throughput\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FPortkey-2025.03-blue\" alt=\"Portkey 徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>企业级 AI 案例研究\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>案例研究：摩根大通的 COiN 平台——用于金融分析的代理式 AI\u003C\u002Fb>\u003C\u002Fi>, AI Mindset Research, \u003Ca href=\"https:\u002F\u002Fwww.ai-mindset.ai\u002Fenterprise-ai-case-studies#JPMorgan\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FBanking-2025.02-green\" alt=\"银行业徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>案例研究：安永在 Microsoft 365 Copilot 中集成的代理式 AI\u003C\u002Fb>\u003C\u002Fi>, AI Mindset Research, \u003Ca href=\"https:\u002F\u002Fwww.ai-mindset.ai\u002Fenterprise-ai-case-studies#EY\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FProfessional_Services-2025.02-green\" alt=\"专业服务行业徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>上下文决定一切：让 AI 真正落地的关键转变\u003C\u002Fb>\u003C\u002Fi>, Phil Mora, \u003Ca href=\"https:\u002F\u002Fwww.philmora.com\u002Fthe-big-picture\u002Fcontext-is-everything-the-massive-shift-making-ai-actually-work-in-the-real-world\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FCross_Industry-2025.06-green\" alt=\"跨行业徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>企业应用与基础设施\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>面向企业 RAG 应用的上下文层\u003C\u002Fb>\u003C\u002Fi>, Contextual AI 团队, \u003Ca href=\"https:\u002F\u002Fcontextual.ai\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FContextual_AI-2025.07-blue\" alt=\"Contextual AI 徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>AI 模型部署的挑战与解决方案\u003C\u002Fb>\u003C\u002Fi>, Dean Lancaster, \u003Ca href=\"https:\u002F\u002Fwww.linkedin.com\u002Fpulse\u002Ffrom-poc-production-overcoming-ai-deployment-ensuring-dean-lancaster-fmtoe\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLinkedIn-2025.03-blue\" alt=\"LinkedIn 徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>2024 年：生成式 AI 在企业中的现状\u003C\u002Fb>\u003C\u002Fi>, Menlo Ventures, \u003Ca href=\"https:\u002F\u002Fmenlovc.com\u002F2024-the-state-of-generative-ai-in-the-enterprise\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FReport-2024-blue\" alt=\"报告徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>100 位企业 CIO 如何在 2025 年构建和采购生成式 AI\u003C\u002Fb>\u003C\u002Fi>, Andreessen Horowitz, \u003Ca href=\"https:\u002F\u002Fa16z.com\u002Fai-enterprise-2025\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fa16z-2025-blue\" alt=\"a16z 徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Cb>结合上下文工程的开发者工具\u003C\u002Fb>\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>Autohand Code CLI：具备语义搜索、记忆与上下文管理功能的自主编码代理\u003C\u002Fb>\u003C\u002Fi>, Autohand AI, \u003Ca href=\"https:\u002F\u002Fwww.autohand.ai\u002Fcode\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FTool-2025-green\" alt=\"工具徽章\">\u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fautohandai\u002Fcode-cli\" target=\"_blank\">\n  \t\t\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fautohandai\u002Fcode-cli.svg?style=social\" alt=\"GitHub 星标\">\n    \u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n#### 编码代理与项目记忆\n\n编码代理是上下文工程演变为代理工程最为清晰的生产应用场景之一。在此，上下文不再仅仅是提示词，而是转变为代码库指令、项目记忆、任务计划、文件差异、测试结果以及工具调用记录。\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>推出 Codex\u003C\u002Fb>\u003C\u002Fi>, OpenAI, \u003Ca href=\"https:\u002F\u002Fopenai.com\u002Findex\u002Fintroducing-codex\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpenAI-2025.05-blue\" alt=\"OpenAI 徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Claude 的代码记忆\u003C\u002Fb>\u003C\u002Fi>, Anthropic, \u003Ca href=\"https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fdocs\u002Fclaude-code\u002Fmemory\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAnthropic-2026-blue\" alt=\"Anthropic 徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Claude 的子代理\u003C\u002Fb>\u003C\u002Fi>, Anthropic, \u003Ca href=\"https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fdocs\u002Fclaude-code\u002Fsub-agents\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAnthropic-2026-blue\" alt=\"Anthropic 徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Letta 的记忆块\u003C\u002Fb>\u003C\u002Fi>, Letta, \u003Ca href=\"https:\u002F\u002Fdocs.letta.com\u002Fguides\u002Fcore-concepts\u002Fmemory\u002Fmemory-blocks\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLetta-2026-blue\" alt=\"Letta 徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LangChain 的深度代理\u003C\u002Fb>\u003C\u002Fi>, LangChain, \u003Ca href=\"https:\u002F\u002Fdocs.langchain.com\u002Foss\u002Fpython\u002Fdeepagents\u002Foverview\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLangChain-2026-blue\" alt=\"LangChain 徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n#### 平台栈与托管代理运行时\n\n如今的生产生态系统正日益围绕完整的代理栈而非孤立的模型或提示词来构建。这些栈将工具、记忆、运行时编排、会话管理、可观测性及互操作性整合于单一平台界面中。\n\n\u003Cul>\n\u003Cli>\u003Ci>\u003Cb>OpenAI 代理指南\u003C\u002Fb>\u003C\u002Fi>, OpenAI, \u003Ca href=\"https:\u002F\u002Fplatform.openai.com\u002Fdocs\u002Fguides\u002Fagents\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpenAI-2026-blue\" alt=\"OpenAI 徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Google 代理开发套件（ADK）\u003C\u002Fb>\u003C\u002Fi>, Google, \u003Ca href=\"https:\u002F\u002Fgoogle.github.io\u002Fadk-docs\u002F\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FGoogle-2026-blue\" alt=\"Google 徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Vertex AI 代理引擎\u003C\u002Fb>\u003C\u002Fi>, Google, \u003Ca href=\"https:\u002F\u002Fcloud.google.com\u002Fvertex-ai\u002Fgenerative-ai\u002Fdocs\u002Fagent-engine\u002Foverview\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FGoogle-2026-blue\" alt=\"Google 徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>LangGraph 记忆概览\u003C\u002Fb>\u003C\u002Fi>, LangChain, \u003Ca href=\"https:\u002F\u002Fdocs.langchain.com\u002Foss\u002Fjavascript\u002Flanggraph\u002Fmemory\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLangChain-2026-blue\" alt=\"LangChain 徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ci>\u003Cb>Microsoft 代理框架\u003C\u002Fb>\u003C\u002Fi>, Microsoft, \u003Ca href=\"https:\u002F\u002Flearn.microsoft.com\u002Fen-us\u002Fagent-framework\u002Fuser-guide\u002Foverview\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FMicrosoft-2026-blue\" alt=\"Microsoft 徽章\">\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\n---\n\n## 🔮 局限性与未来方向\n\n### 当前局限性\n\n1. **上下文窗口限制**：尽管有所改进，但上下文长度仍然是瓶颈\n2. **计算开销**：处理大规模上下文需要大量资源\n3. **上下文连贯性**：在扩展的上下文中保持一致性仍具挑战\n4. **动态适应性**：实时更新上下文面临困难\n\n### 未来研究方向\n\n1. **无限上下文**：开发真正无限制的上下文处理能力\n2. **上下文压缩**：高效表示大规模上下文\n3. **多模态融合**：无缝整合多种数据类型\n4. **自适应上下文**：自我优化的上下文管理\n5. **上下文隐私**：保障上下文处理流程中的敏感信息安全\n\n---\n\n## 🤝 贡献\n\n我们欢迎对本综述的贡献！请遵循以下指南：\n\n1. **Fork** 该仓库\n2. **创建** 功能分支\n3. **添加** 相关论文，并确保格式正确\n4. **提交** 拉取请求，附上清晰的描述\n\n### 论文格式指南\n\n```markdown\n\u003Cli>\u003Ci>\u003Cb>论文标题\u003C\u002Fb>\u003C\u002Fi>, 作者等, \u003Ca href=\"URL\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FSOURCE-YEAR.MM-COLOR\" alt=\"SOURCE Badge\">\u003C\u002Fa>\u003C\u002Fli>\n```\n\n### 徽章颜色\n- ![arXiv徽章](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-red) `red` 用于 arXiv 论文\n- ![PDF徽章](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FPDF-blue) `blue` 用于会议\u002F期刊论文\n- ![GitHub徽章](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FGitHub-white) `white` 用于 GitHub 仓库\n- ![HuggingFace徽章](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FHuggingFace-yellow) `yellow` 用于 HuggingFace 资源\n\n---\n\n## 📄 许可证\n\n本项目采用 MIT 许可证授权——详情请参阅 [LICENSE](LICENSE) 文件。\n\n---\n\n## 📑 引用\n\n如果您在研究中发现本综述有所帮助，请考虑引用：\n\n```bibtex\n@misc{mei2025surveycontextengineeringlarge,\n      title={大型语言模型上下文工程综述}, \n      author={Mei Lingrui, Yao Jiayu, Ge Yuyao, Wang Yiwei, Bi Baolong, Cai Yujun, Liu Jiazhi, Li Mingyu, Li Zhong-Zhi, Zhang Duzhen, Zhou Chenlin, Mao Jiayi, Xia Tianze, Guo Jiafeng, Liu Shenghua},\n      year={2025},\n      eprint={2507.13334},\n      archivePrefix={arXiv},\n      primaryClass={cs.CL},\n      url={https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.13334}, \n}\n```\n\n---\n\n## ⚠️ 免责声明\n\n本项目处于**持续更新**和不断发展之中。尽管我们力求准确和全面，但仍可能存在错误、遗漏或过时的信息。我们欢迎社区成员提出修正意见、建议及贡献。请持续关注我们的定期更新与改进。\n\n## 📧 联系方式\n\n如有任何问题、建议或合作机会，请随时联系我们：\n\n**Mei Lingrui**  \n📧 邮箱：[meilingrui22@mails.ucas.ac.cn](mailto:meilingrui22@mails.ucas.ac.cn)\n\n您也可以在此仓库中提交议题，进行一般性讨论和建议。\n\n---\n\n## 🙏 致谢\n\n本综述建立在人工智能研究社区的奠基性工作之上。我们感谢所有为上下文工程及大型语言模型发展做出贡献的研究人员。\n\n---\n\n## 星标历史\n\n**如果您觉得本项目有帮助，请为它点亮星星⭐！**\n\n[![星标历史图](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FMeirtz_Awesome-Context-Engineering_readme_bc8c8bff666e.png)](https:\u002F\u002Fwww.star-history.com\u002F#Meirtz\u002FAwesome-Context-Engineering&Date)\n\n---\n\n## 📖 我们的论文\n\n**大型语言模型上下文工程综述**\n\n- **arXiv**: https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.13334\n- **Hugging Face Papers**: https:\u002F\u002Fhuggingface.co\u002Fpapers\u002F2507.13334\n\n这篇综合性的综述提供了关于大型语言模型上下文工程的最新学术见解和理论基础。","# Awesome-Context Engineering 快速上手指南\n\n**Awesome-Context Engineering** 并非一个可直接安装的软件包或库，而是一个**综合性的资源清单（Awesome List）和学术综述项目**。它汇集了关于“上下文工程（Context Engineering）”从静态提示词到动态代理系统（Agent Systems）、记忆系统、协议及可观测性栈的最新论文、工具、博客和技术架构。\n\n本指南旨在帮助开发者快速利用该仓库提供的资源，构建生产级的 AI 应用。\n\n## 环境准备\n\n由于本项目是资源索引，无需特定的运行时环境，但为了实践其中的技术（如 RAG、Agent 开发、上下文管理），建议准备以下基础开发环境：\n\n*   **操作系统**: Linux, macOS 或 Windows (WSL2 推荐)\n*   **编程语言**: Python 3.9+ (AI 生态主流语言)\n*   **包管理器**: `pip` 或 `conda`\n*   **版本控制**: Git\n*   **前置依赖库** (根据你选择的具体技术栈安装，通用推荐):\n    ```bash\n    pip install langchain langgraph openai anthropic google-generativeai\n    pip install llama-index chromadb qdrant-client\n    pip install pydantic httpx\n    ```\n*   **国内加速方案**:\n    *   使用清华或阿里镜像源加速 Python 包安装：\n        ```bash\n        pip install -i https:\u002F\u002Fpypi.tuna.tsinghua.edu.cn\u002Fsimple \u003Cpackage_name>\n        ```\n    *   若访问 GitHub 或 arXiv 受限，建议使用学术加速工具或配置国内镜像代理。\n\n## 安装步骤\n\n你需要克隆该仓库以获取最新的资源列表、综述论文草稿及分类索引。\n\n1.  **克隆仓库**:\n    ```bash\n    git clone https:\u002F\u002Fgithub.com\u002FMeilingrui\u002FAwesome-Context-Engineering.git\n    ```\n\n2.  **进入目录**:\n    ```bash\n    cd Awesome-Context-Engineering\n    ```\n\n3.  **查看资源结构**:\n    直接打开 `README.md` 文件，或在终端查看目录结构（如果仓库包含代码示例文件夹）：\n    ```bash\n    ls -la\n    cat README.md\n    ```\n\n4.  **获取核心论文**:\n    该项目关联的综述论文已发布，可直接通过以下链接获取最新理论框架：\n    *   arXiv: [A Survey of Context Engineering for Large Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.13334)\n    *   Hugging Face Papers: [Link](https:\u002F\u002Fhuggingface.co\u002Fpapers\u002F2507.13334)\n\n## 基本使用\n\n本项目的核心用法是**查阅索引**并**应用推荐的技术栈**。以下是基于仓库内容的典型使用流程：\n\n### 1. 查找特定技术领域的资源\n假设你需要构建一个具备长期记忆的编码助手（Coding Agent with Project Memory），请在 `README.md` 中定位到 **\"🚀 Applications and Systems\"** 或 **\"🛠️ Implementation and Challenges\"** 章节。\n\n*   **搜索关键词**: 在文件中搜索 `Memory Systems`, `Coding Agents`, 或 `MCP` (Model Context Protocol)。\n*   **参考案例**: 仓库推荐了以下具体实现方案：\n    *   **记忆管理**: 参考 [Letta memory blocks](https:\u002F\u002Fdocs.letta.com\u002Fguides\u002Fcore-concepts\u002Fmemory\u002Fmemory-blocks) 或 [Claude Code memory](https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fdocs\u002Fclaude-code\u002Fmemory)。\n    *   **协议标准**: 查阅 [Model Context Protocol (MCP)](https:\u002F\u002Fmodelcontextprotocol.io\u002Fspecification\u002F2025-06-18) 以实现工具互操作性。\n\n### 2. 构建动态上下文系统 (示例代码逻辑)\n根据仓库中 **\"Dynamic Context Orchestration\"** 的理念，不要仅使用静态 Prompt，而是构建动态上下文加载器。以下是一个基于 LangChain 的简化概念示例，体现“上下文工程”的核心思想：\n\n```python\nfrom langchain_core.prompts import ChatPromptTemplate\nfrom langchain_core.runnables import RunnablePassthrough\n\n# 1. 定义动态上下文组件 (而非单一字符串)\nsystem_context = \"You are an expert coding assistant with access to project memory.\"\nretrieved_knowledge = \"{context_from_rag}\"\nuser_instruction = \"{question}\"\n\n# 2. 构建结构化上下文模板\nprompt_template = ChatPromptTemplate.from_messages([\n    (\"system\", system_context),\n    (\"human\", f\"Current Project State:\\n{retrieved_knowledge}\\n\\nUser Task: {user_instruction}\")\n])\n\n# 3. 动态编排 (Orchestration)\n# 在实际生产中，这里会接入向量数据库检索、历史对话压缩、工具调用结果等\nchain = prompt_template | model\n\n# 4. 执行\nresponse = chain.invoke({\n    \"context_from_rag\": \"import pandas as pd...\", # 来自 RAG 检索的动态内容\n    \"question\": \"Refactor this data loading function.\"\n})\n```\n\n### 3. 跟进 2026 代理时代新特性\n仓库特别更新了 **\"2026 Agent Era Update\"** 章节，重点关注以下方向：\n*   **Agent Harnesses**: 学习如何管理子代理、检查点（Checkpoints）和沙箱环境。参考 [Anthropic Effective Agents Guide](https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Fbuilding-effective-agents)。\n*   **可观测性 (Observability)**: 集成 [LangSmith](https:\u002F\u002Fdocs.langchain.com\u002Flangsmith\u002Fobservability-quickstart) 或 [OpenTelemetry](https:\u002F\u002Fopentelemetry.io\u002Fdocs\u002Fspecs\u002Fsemconv\u002Fgen-ai\u002F) 来监控长运行代理系统的状态。\n*   **开放协议**: 研究 [A2A](https:\u002F\u002Fa2a-protocol.org\u002Flatest\u002F) 和 [AgentSchema](https:\u002F\u002Fmicrosoft.github.io\u002FAgentSchema\u002F) 以实现不同代理间的通信。\n\n### 4. 参与社区与贡献\n*   **加入讨论**: 扫描仓库中的二维码加入微信群，或访问 [Discord 服务器](https:\u002F\u002Fdiscord.gg\u002Ffsqs3Ybh)。\n*   **提交资源**: 如果你发现新的上下文工程工具或论文，可以通过 GitHub Issue 或 Pull Request 提交到该仓库。\n*   **联系作者**: 如有合作意向，可通过 `meilingrui25b@ict.ac.cn` 联系。\n\n通过遵循上述步骤，你可以充分利用 **Awesome-Context Engineering** 提供的知识体系，从传统的提示词工程进阶到构建复杂、可靠的生产级 AI 代理系统。","某金融科技公司正在构建一个需要处理长周期任务、具备记忆能力且能调用外部工具的复杂 AI 客服代理系统。\n\n### 没有 Awesome-Context-Engineering 时\n- **架构设计盲目**：团队仅依赖静态提示词，缺乏对动态上下文管理、记忆系统及代理运行时状态的理论指导，导致系统在多轮对话中频繁“失忆”。\n- **资源筛选低效**：面对海量零散的论文和框架，开发人员难以辨别哪些技术适合生产环境，花费数周时间试错却找不到成熟的实现指南。\n- **可观测性缺失**：系统上线后无法有效追踪代理的决策路径和上下文压缩效果，出现错误时只能靠猜测排查，运维成本极高。\n- **扩展性受限**：由于未掌握代理间通信协议和工具调用规范，每当新增业务功能（如审批流程或沙箱执行），都需要重构核心代码。\n\n### 使用 Awesome-Context-Engineering 后\n- **架构清晰落地**：直接参考其收录的生产级代理运行时方案和记忆工件格式，快速构建了支持长程任务规划和人工审批循环的稳健架构。\n- **技术选型精准**：利用其整理的数百篇前沿论文和框架清单，团队迅速锁定了最适合金融场景的上下文压缩与缓存策略，研发周期缩短 60%。\n- **监控体系完善**：依据其推荐的“追踪优先”可观测性栈，实现了从提示词加载到代理决策的全链路监控，故障定位时间从小时级降至分钟级。\n- **平滑演进升级**：借助其对 2026 代理时代互操作协议的梳理，系统轻松集成了新的编码代理和项目记忆模块，无需推翻重来即可支持复杂业务扩展。\n\nAwesome-Context-Engineering 将团队从碎片化的提示词调试中解放出来，提供了通往生产级智能代理系统的完整工程化地图。","https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FMeirtz_Awesome-Context-Engineering_f88855f1.png","Meirtz","Lingrui Mei","https:\u002F\u002Foss.gittoolsai.com\u002Favatars\u002FMeirtz_585ea4a8.jpg","PhD Student at Chinese Academy of Sciences. \r\nIntern at Tencent HY.\r\nComputational Linguistics \u002F Reinforcement Learning \u002F Racing \u002F Gaming","Chinese Academy of Sciences","Beijing, China","meilingrui22@mails.ucas.ac.cn",null,"https:\u002F\u002Fme.meirtz.com\u002Fabout","https:\u002F\u002Fgithub.com\u002FMeirtz",3067,214,"2026-04-18T00:29:52","MIT","","未说明",{"notes":91,"python":89,"dependencies":92},"该项目是一个资源列表（Awesome List）和综述论文集合，而非可执行的软件工具或代码库。它主要收集了关于上下文工程、智能体运行时、记忆系统、协议和可观测性栈的相关资源、论文和技术链接。因此，该项目本身没有特定的操作系统、GPU、内存、Python 版本或依赖库要求。用户只需浏览 README 中列出的外部链接即可获取相关信息。",[],[13,14,94,35],"其他",[96,97,98,99,100,101,102,103],"agent","agentic-ai","agi","awesome-list","cognitive-science","context-engineering","llm","rag","2026-03-27T02:49:30.150509","2026-04-19T06:02:08.527353",[],[]]