[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"similar-elicit--machine-learning-list":3,"tool-elicit--machine-learning-list":61},[4,18,26,36,44,53],{"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",143909,2,"2026-04-07T11:33:18",[14,13,35],"语言模型",{"id":37,"name":38,"github_repo":39,"description_zh":40,"stars":41,"difficulty_score":32,"last_commit_at":42,"category_tags":43,"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。",107888,"2026-04-06T11:32:50",[14,15,13],{"id":45,"name":46,"github_repo":47,"description_zh":48,"stars":49,"difficulty_score":32,"last_commit_at":50,"category_tags":51,"status":17},4721,"markitdown","microsoft\u002Fmarkitdown","MarkItDown 是一款由微软 AutoGen 团队打造的轻量级 Python 工具，专为将各类文件高效转换为 Markdown 格式而设计。它支持 PDF、Word、Excel、PPT、图片（含 OCR）、音频（含语音转录）、HTML 乃至 YouTube 链接等多种格式的解析，能够精准提取文档中的标题、列表、表格和链接等关键结构信息。\n\n在人工智能应用日益普及的今天，大语言模型（LLM）虽擅长处理文本，却难以直接读取复杂的二进制办公文档。MarkItDown 恰好解决了这一痛点，它将非结构化或半结构化的文件转化为模型“原生理解”且 Token 效率极高的 Markdown 格式，成为连接本地文件与 AI 分析 pipeline 的理想桥梁。此外，它还提供了 MCP（模型上下文协议）服务器，可无缝集成到 Claude Desktop 等 LLM 应用中。\n\n这款工具特别适合开发者、数据科学家及 AI 研究人员使用，尤其是那些需要构建文档检索增强生成（RAG）系统、进行批量文本分析或希望让 AI 助手直接“阅读”本地文件的用户。虽然生成的内容也具备一定可读性，但其核心优势在于为机器",93400,"2026-04-06T19:52:38",[52,14],"插件",{"id":54,"name":55,"github_repo":56,"description_zh":57,"stars":58,"difficulty_score":10,"last_commit_at":59,"category_tags":60,"status":17},4487,"LLMs-from-scratch","rasbt\u002FLLMs-from-scratch","LLMs-from-scratch 是一个基于 PyTorch 的开源教育项目，旨在引导用户从零开始一步步构建一个类似 ChatGPT 的大型语言模型（LLM）。它不仅是同名技术著作的官方代码库，更提供了一套完整的实践方案，涵盖模型开发、预训练及微调的全过程。\n\n该项目主要解决了大模型领域“黑盒化”的学习痛点。许多开发者虽能调用现成模型，却难以深入理解其内部架构与训练机制。通过亲手编写每一行核心代码，用户能够透彻掌握 Transformer 架构、注意力机制等关键原理，从而真正理解大模型是如何“思考”的。此外，项目还包含了加载大型预训练权重进行微调的代码，帮助用户将理论知识延伸至实际应用。\n\nLLMs-from-scratch 特别适合希望深入底层原理的 AI 开发者、研究人员以及计算机专业的学生。对于不满足于仅使用 API，而是渴望探究模型构建细节的技术人员而言，这是极佳的学习资源。其独特的技术亮点在于“循序渐进”的教学设计：将复杂的系统工程拆解为清晰的步骤，配合详细的图表与示例，让构建一个虽小但功能完备的大模型变得触手可及。无论你是想夯实理论基础，还是为未来研发更大规模的模型做准备",90106,"2026-04-06T11:19:32",[35,15,13,14],{"id":62,"github_repo":63,"name":64,"description_en":65,"description_zh":66,"ai_summary_zh":67,"readme_en":68,"readme_zh":69,"quickstart_zh":70,"use_case_zh":71,"hero_image_url":72,"owner_login":73,"owner_name":74,"owner_avatar_url":75,"owner_bio":76,"owner_company":77,"owner_location":77,"owner_email":77,"owner_twitter":78,"owner_website":79,"owner_url":80,"languages":77,"stars":81,"forks":82,"last_commit_at":83,"license":77,"difficulty_score":84,"env_os":85,"env_gpu":86,"env_ram":86,"env_deps":87,"category_tags":90,"github_topics":91,"view_count":32,"oss_zip_url":77,"oss_zip_packed_at":77,"status":17,"created_at":96,"updated_at":97,"faqs":98,"releases":99},5018,"elicit\u002Fmachine-learning-list","machine-learning-list","A curriculum for learning about foundation models, from scratch to the frontier","machine-learning-list 是一份专为系统掌握基础模型（Foundation Models）而设计的开源学习大纲，内容涵盖从零基础入门到前沿技术探索的全路径。它旨在解决机器学习领域知识更新快、资料分散且难以构建完整认知体系的痛点，帮助学习者高效筛选出最具价值的核心论文与教程。\n\n这份清单最初用于指导 Elicit 团队的新员工快速建立机器学习背景，特别聚焦于语言模型。其内容结构严谨，按“基础理论、推理策略、实际应用、工程实践、进阶主题及宏观视野”六大板块组织，并创新性地采用“分级阅读”模式：用户可先攻克 Tier 1 核心概念，再逐步深入 Tier 2 及更高阶内容。资源形式丰富，既包含经典的学术文章，也精选了 Karpathy 等专家的高质量视频讲解，兼顾生产部署技巧与长期可扩展性技术。\n\nmachine-learning-list 非常适合希望转行或深耕 AI 领域的开发者、研究人员，以及需要快速补齐大模型知识短板的工程师使用。无论你是想理解 Transformer 架构底层原理，还是探索 AI 安全、世界模型等前沿议题，都能在这里找到清晰的学习指引。它不仅是一份书","machine-learning-list 是一份专为系统掌握基础模型（Foundation Models）而设计的开源学习大纲，内容涵盖从零基础入门到前沿技术探索的全路径。它旨在解决机器学习领域知识更新快、资料分散且难以构建完整认知体系的痛点，帮助学习者高效筛选出最具价值的核心论文与教程。\n\n这份清单最初用于指导 Elicit 团队的新员工快速建立机器学习背景，特别聚焦于语言模型。其内容结构严谨，按“基础理论、推理策略、实际应用、工程实践、进阶主题及宏观视野”六大板块组织，并创新性地采用“分级阅读”模式：用户可先攻克 Tier 1 核心概念，再逐步深入 Tier 2 及更高阶内容。资源形式丰富，既包含经典的学术文章，也精选了 Karpathy 等专家的高质量视频讲解，兼顾生产部署技巧与长期可扩展性技术。\n\nmachine-learning-list 非常适合希望转行或深耕 AI 领域的开发者、研究人员，以及需要快速补齐大模型知识短板的工程师使用。无论你是想理解 Transformer 架构底层原理，还是探索 AI 安全、世界模型等前沿议题，都能在这里找到清晰的学习指引。它不仅是一份书单，更是一张通往大模型技术深处的可靠地图。","# Elicit Machine Learning Reading List\n\n## Purpose\n\nThe purpose of this curriculum is to help new [Elicit](https:\u002F\u002Felicit.com\u002F) employees learn background in machine learning, with a focus on language models. I’ve tried to strike a balance between papers that are relevant for deploying ML in production and techniques that matter for longer-term scalability.\n\nIf you don’t work at Elicit yet - we’re [hiring ML and software engineers](https:\u002F\u002Felicit.com\u002Fcareers).\n\n## How to read\n\nRecommended reading order:\n\n1. Read “Tier 1” for all topics\n2. Read “Tier 2” for all topics\n3. Etc\n\n✨ = Added after 2025\u002F11\u002F26\n\n## Table of contents\n\n- [Fundamentals](#fundamentals)\n  - [Introduction to machine learning](#introduction-to-machine-learning)\n  - [Transformers](#transformers)\n  - [Key foundation model architectures](#key-foundation-model-architectures)\n  - [Training and finetuning](#training-and-finetuning)\n- [Reasoning and runtime strategies](#reasoning-and-runtime-strategies)\n  - [In-context reasoning](#in-context-reasoning)\n  - [Task decomposition](#task-decomposition)\n  - [Debate](#debate)\n  - [Tool use and scaffolding](#tool-use-and-scaffolding)\n  - [Honesty, factuality, and epistemics](#honesty-factuality-and-epistemics)\n- [Applications](#applications)\n  - [Science](#science)\n  - [Forecasting](#forecasting)\n  - [Search and ranking](#search-and-ranking)\n- [ML in practice](#ml-in-practice)\n  - [Production deployment](#production-deployment)\n  - [Benchmarks](#benchmarks)\n  - [Datasets](#datasets)\n- [Advanced topics](#advanced-topics)\n  - [World models and causality](#world-models-and-causality)\n  - [Planning](#planning)\n  - [Uncertainty, calibration, and active learning](#uncertainty-calibration-and-active-learning)\n  - [Interpretability and model editing](#interpretability-and-model-editing)\n  - [Reinforcement learning](#reinforcement-learning)\n- [The big picture](#the-big-picture)\n  - [AI scaling](#ai-scaling)\n  - [AI safety](#ai-safety)\n  - [Economic and social impacts](#economic-and-social-impacts)\n  - [Philosophy](#philosophy)\n- [Maintainer](#maintainer)\n\n## Fundamentals\n\n### Introduction to machine learning\n\n**Tier 1**\n\n- [A short introduction to machine learning](https:\u002F\u002Fwww.alignmentforum.org\u002Fposts\u002FqE73pqxAZmeACsAdF\u002Fa-short-introduction-to-machine-learning)\n- [But what is a neural network?](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=aircAruvnKk&t=0s)\n- [Gradient descent, how neural networks learn](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=IHZwWFHWa-w)\n\n**Tier 2**\n\n- [An intuitive understanding of backpropagation](https:\u002F\u002Fcs231n.github.io\u002Foptimization-2\u002F)\n- [What is backpropagation really doing?](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=Ilg3gGewQ5U&list=PLZHQObOWTQDNU6R1_67000Dx_ZCJB-3pi&index=4)\n- [An introduction to deep reinforcement learning](https:\u002F\u002Fthomassimonini.medium.com\u002Fan-introduction-to-deep-reinforcement-learning-17a565999c0c)\n\n**Tier 3**\n\n- [The spelled-out intro to neural networks and backpropagation: building micrograd](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=VMj-3S1tku0) (Karpathy)\n- [Backpropagation calculus](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=tIeHLnjs5U8&list=PLZHQObOWTQDNU6R1_67000Dx_ZCJB-3pi&index=5)\n\n### Transformers\n\n**Tier 1**\n\n- ✨ [Intro to Large Language Models](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=zjkBMFhNj_g) (Karpathy)\n- [But what is a GPT? Visual intro to transformers](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=wjZofJX0v4M)\n- [Attention in transformers, visually explained](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=eMlx5fFNoYc)\n- [Attention? Attention!](https:\u002F\u002Flilianweng.github.io\u002Fposts\u002F2018-06-24-attention\u002F)\n- [The Illustrated Transformer](http:\u002F\u002Fjalammar.github.io\u002Fillustrated-transformer\u002F)\n\n**Tier 2**\n\n- ✨ [Deep Dive into LLMs like ChatGPT](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=7xTGNNLPyMI) (Karpathy)\n- [Let's build the GPT Tokenizer](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=zduSFxRajkE) (Karpathy)\n- [The Illustrated GPT-2 (Visualizing Transformer Language Models)](https:\u002F\u002Fjalammar.github.io\u002Fillustrated-gpt2\u002F)\n- [Neural Machine Translation by Jointly Learning to Align and Translate](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1409.0473)\n- [Attention Is All You Need](https:\u002F\u002Farxiv.org\u002Fabs\u002F1706.03762)\n\n**Tier 3**\n\n- ✨ [The Reversal Curse: LLMs trained on \"A is B\" fail to learn \"B is A\"](https:\u002F\u002Farxiv.org\u002Fabs\u002F2309.12288)\n- [The Annotated Transformer](https:\u002F\u002Fnlp.seas.harvard.edu\u002F2018\u002F04\u002F03\u002Fattention.html)\n- [TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second](https:\u002F\u002Farxiv.org\u002Fabs\u002F2207.01848)\n- [Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets](https:\u002F\u002Farxiv.org\u002Fabs\u002F2201.02177)\n- [A Mathematical Framework for Transformer Circuits](https:\u002F\u002Ftransformer-circuits.pub\u002F2021\u002Fframework\u002Findex.html)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>Tier 4+\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [A Practical Survey on Faster and Lighter Transformers](https:\u002F\u002Farxiv.org\u002Fabs\u002F2103.14636)\n- [Compositional Capabilities of Autoregressive Transformers: A Study on Synthetic, Interpretable Tasks](https:\u002F\u002Farxiv.org\u002Fabs\u002F2311.12997)\n- [Memorizing Transformers](https:\u002F\u002Farxiv.org\u002Fabs\u002F2203.08913)\n- [Transformer Feed-Forward Layers Are Key-Value Memories](https:\u002F\u002Farxiv.org\u002Fabs\u002F2012.14913)\n\n\u003C\u002Fdetails>\n\n### Key foundation model architectures\n\n**Tier 1**\n\n- [Language Models are Unsupervised Multitask Learners](https:\u002F\u002Fwww.semanticscholar.org\u002Fpaper\u002FLanguage-Models-are-Unsupervised-Multitask-Learners-Radford-Wu\u002F9405cc0d6169988371b2755e573cc28650d14dfe) (GPT-2)\n- [Language Models are Few-Shot Learners](https:\u002F\u002Farxiv.org\u002Fabs\u002F2005.14165) (GPT-3)\n\n**Tier 2**\n\n- ✨ [DeepSeek-R1](https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.12948) (DeepSeek-R1)\n- ✨ [DeepSeek-V3 Technical Report](https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.19437) (DeepSeek-V3)\n- ✨ [The Llama 3 Herd of Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.21783) (Llama 3)\n- [LLaMA: Open and Efficient Foundation Language Models](http:\u002F\u002Farxiv.org\u002Fabs\u002F2302.13971) (LLaMA)\n- [Training language models to follow instructions with human feedback](https:\u002F\u002Farxiv.org\u002Fabs\u002F2203.02155) (OpenAI Instruct)\n\n**Tier 3**\n\n- ✨ [LLaMA 2: Open Foundation and Fine-Tuned Chat Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2307.09288) (LLaMA 2)\n- ✨ [Qwen2.5 Technical Report](https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.15115) (Qwen2.5)\n- ✨ [Titans: Learning to Memorize at Test Time](https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.00663)\n- ✨ [Byte Latent Transformer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.09871)\n- ✨ [Phi-4 Technical Report](https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.08905) (phi-4)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>Tier 4+\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [Evaluating Large Language Models Trained on Code](https:\u002F\u002Farxiv.org\u002Fabs\u002F2107.03374) (OpenAI Codex)\n- [Mistral 7B](http:\u002F\u002Farxiv.org\u002Fabs\u002F2310.06825) (Mistral)\n- [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https:\u002F\u002Farxiv.org\u002Fabs\u002F1910.10683) (T5)\n- [Gemini: A Family of Highly Capable Multimodal Models](https:\u002F\u002Fstorage.googleapis.com\u002Fdeepmind-media\u002Fgemini\u002Fgemini_1_report.pdf) (Gemini)\n- [Mamba: Linear-Time Sequence Modeling with Selective State Spaces](https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.00752v1) (Mamba)\n- [Scaling Instruction-Finetuned Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.11416) (Flan)\n- [Efficiently Modeling Long Sequences with Structured State Spaces](https:\u002F\u002Farxiv.org\u002Fabs\u002F2111.00396) ([video](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=EvQ3ncuriCM)) (S4)\n- [Consistency Models](http:\u002F\u002Farxiv.org\u002Fabs\u002F2303.01469)\n- [Model Card and Evaluations for Claude Models](https:\u002F\u002Fwww-cdn.anthropic.com\u002Fbd2a28d2535bfb0494cc8e2a3bf135d2e7523226\u002FModel-Card-Claude-2.pdf) (Claude 2)\n- [OLMo: Accelerating the Science of Language Models](http:\u002F\u002Farxiv.org\u002Fabs\u002F2402.00838)\n- [PaLM 2 Technical Report](https:\u002F\u002Farxiv.org\u002Fabs\u002F2305.10403) (Palm 2)\n- [Textbooks Are All You Need II: phi-1.5 technical report](http:\u002F\u002Farxiv.org\u002Fabs\u002F2309.05463) (phi 1.5)\n- [Visual Instruction Tuning](http:\u002F\u002Farxiv.org\u002Fabs\u002F2304.08485) (LLaVA)\n- [A General Language Assistant as a Laboratory for Alignment](https:\u002F\u002Farxiv.org\u002Fabs\u002F2112.00861)\n- [Finetuned Language Models Are Zero-Shot Learners](https:\u002F\u002Farxiv.org\u002Fabs\u002F2109.01652) (Google Instruct)\n- [Galactica: A Large Language Model for Science](https:\u002F\u002Farxiv.org\u002Fabs\u002F2211.09085)\n- [LaMDA: Language Models for Dialog Applications](https:\u002F\u002Farxiv.org\u002Fabs\u002F2201.08239) (Google Dialog)\n- [OPT: Open Pre-trained Transformer Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2112.11446) (Meta GPT-3)\n- [PaLM: Scaling Language Modeling with Pathways](https:\u002F\u002Farxiv.org\u002Fabs\u002F2204.02311) (PaLM)\n- [Program Synthesis with Large Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2108.07732) (Google Codex)\n- [Scaling Language Models: Methods, Analysis & Insights from Training Gopher](https:\u002F\u002Farxiv.org\u002Fabs\u002F2112.11446) (Gopher)\n- [Solving Quantitative Reasoning Problems with Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2206.14858) (Minerva)\n- [UL2: Unifying Language Learning Paradigms](http:\u002F\u002Faima.cs.berkeley.edu\u002F) (UL2)\n\n\u003C\u002Fdetails>\n\n### Training and finetuning\n\n**Tier 2**\n\n- ✨ [Deep Learning Tuning Playbook](https:\u002F\u002Fgithub.com\u002Fgoogle-research\u002Ftuning_playbook)\n- [Tensor Programs V: Tuning Large Neural Networks via Zero-Shot Hyperparameter Transfer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2203.03466)\n- [Learning to summarise with human feedback](https:\u002F\u002Farxiv.org\u002Fabs\u002F2009.01325)\n- [Training Verifiers to Solve Math Word Problems](https:\u002F\u002Farxiv.org\u002Fabs\u002F2110.14168)\n\n**Tier 3**\n\n- ✨ [Better & Faster Large Language Models via Multi-token Prediction](https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.19737)\n- ✨ [LoRA vs Full Fine-tuning: An Illusion of Equivalence](https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.21228)\n- ✨ [QLoRA: Efficient Finetuning of Quantized LLMs](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.14314)\n- [Pretraining Language Models with Human Preferences](http:\u002F\u002Farxiv.org\u002Fabs\u002F2302.08582)\n- [Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision](http:\u002F\u002Farxiv.org\u002Fabs\u002F2312.09390)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>Tier 4+\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [Few-Shot Parameter-Efficient Fine-Tuning is Better and Cheaper than In-Context Learning](https:\u002F\u002Farxiv.org\u002Fabs\u002F2205.05638v1)\n- [Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Models](http:\u002F\u002Farxiv.org\u002Fabs\u002F2312.06585)\n- [Improving Code Generation by Training with Natural Language Feedback](http:\u002F\u002Farxiv.org\u002Fabs\u002F2303.16749)\n- [Language Modeling Is Compression](https:\u002F\u002Farxiv.org\u002Fabs\u002F2309.10668v1)\n- [LIMA: Less Is More for Alignment](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.11206)\n- [Learning to Compress Prompts with Gist Tokens](http:\u002F\u002Farxiv.org\u002Fabs\u002F2304.08467)\n- [Lost in the Middle: How Language Models Use Long Contexts](http:\u002F\u002Farxiv.org\u002Fabs\u002F2307.03172)\n- [LoRA: Low-Rank Adaptation of Large Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2106.09685)\n- [Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking](http:\u002F\u002Farxiv.org\u002Fabs\u002F2403.09629)\n- [Reinforced Self-Training (ReST) for Language Modeling](http:\u002F\u002Farxiv.org\u002Fabs\u002F2308.08998)\n- [Solving olympiad geometry without human demonstrations](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41586-023-06747-5)\n- [Tell, don't show: Declarative facts influence how LLMs generalize](http:\u002F\u002Farxiv.org\u002Fabs\u002F2312.07779)\n- [Textbooks Are All You Need](http:\u002F\u002Farxiv.org\u002Fabs\u002F2306.11644)\n- [TinyStories: How Small Can Language Models Be and Still Speak Coherent English?](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.07759)\n- [Training Language Models with Language Feedback at Scale](http:\u002F\u002Farxiv.org\u002Fabs\u002F2303.16755)\n- [Turing Complete Transformers: Two Transformers Are More Powerful Than One](https:\u002F\u002Fopenreview.net\u002Fforum?id=MGWsPGogLH)\n- [ByT5: Towards a token-free future with pre-trained byte-to-byte models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2105.13626)\n- [Data Distributional Properties Drive Emergent In-Context Learning in Transformers](https:\u002F\u002Farxiv.org\u002Fabs\u002F2205.05055)\n- [Diffusion-LM Improves Controllable Text Generation](https:\u002F\u002Farxiv.org\u002Fabs\u002F2205.14217)\n- [ERNIE 3.0: Large-scale Knowledge Enhanced Pre-training for Language Understanding and Generation](https:\u002F\u002Farxiv.org\u002Fabs\u002F2107.02137)\n- [Efficient Training of Language Models to Fill in the Middle](https:\u002F\u002Farxiv.org\u002Fabs\u002F2207.14255)\n- [ExT5: Towards Extreme Multi-Task Scaling for Transfer Learning](https:\u002F\u002Farxiv.org\u002Fabs\u002F2111.10952)\n- [Prefix-Tuning: Optimizing Continuous Prompts for Generation](https:\u002F\u002Farxiv.org\u002Fabs\u002F2101.00190)\n- [Self-Attention Between Datapoints: Going Beyond Individual Input-Output Pairs in Deep Learning](https:\u002F\u002Farxiv.org\u002Fabs\u002F2106.02584)\n- [True Few-Shot Learning with Prompts -- A Real-World Perspective](https:\u002F\u002Farxiv.org\u002Fabs\u002F2111.13440)\n\n\u003C\u002Fdetails>\n\n## Reasoning and runtime strategies\n\n### In-context reasoning\n\n**Tier 2**\n\n- ✨ [Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters](https:\u002F\u002Farxiv.org\u002Fabs\u002F2408.03314)\n- [Chain of Thought Prompting Elicits Reasoning in Large Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2201.11903)\n- [Large Language Models are Zero-Shot Reasoners](https:\u002F\u002Farxiv.org\u002Fabs\u002F2205.11916) (Let's think step by step)\n- [Self-Consistency Improves Chain of Thought Reasoning in Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2203.11171)\n\n**Tier 3**\n\n- ✨ [s1: Simple test-time scaling](https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.19393)\n- ✨ [Cognitive Behaviors that Enable Self-Improving Reasoners, or, Four Habits of Highly Effective STaRs](https:\u002F\u002Farxiv.org\u002Fabs\u002F2503.01307)\n- ✨ [The Surprising Effectiveness of Test-Time Training for Abstract Reasoning](https:\u002F\u002Farxiv.org\u002Fabs\u002F2411.07279)\n- ✨ [Large Language Models Cannot Self-Correct Reasoning Yet](https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.01798v1)\n- [Chain-of-Thought Reasoning Without Prompting](http:\u002F\u002Farxiv.org\u002Fabs\u002F2402.10200)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>Tier 4+\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [Why think step-by-step? Reasoning emerges from the locality of experience](http:\u002F\u002Farxiv.org\u002Fabs\u002F2304.03843)\n- [Baldur: Whole-Proof Generation and Repair with Large Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2303.04910v1)\n- [Bias-Augmented Consistency Training Reduces Biased Reasoning in Chain-of-Thought](http:\u002F\u002Farxiv.org\u002Fabs\u002F2403.05518)\n- [Certified Reasoning with Language Models](http:\u002F\u002Farxiv.org\u002Fabs\u002F2306.04031)\n- [Hypothesis Search: Inductive Reasoning with Language Models](http:\u002F\u002Farxiv.org\u002Fabs\u002F2309.05660)\n- [LLMs and the Abstraction and Reasoning Corpus: Successes, Failures, and the Importance of Object-based Representations](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.18354)\n- [Stream of Search (SoS): Learning to Search in Language](http:\u002F\u002Farxiv.org\u002Fabs\u002F2404.03683)\n- [Training Chain-of-Thought via Latent-Variable Inference](http:\u002F\u002Farxiv.org\u002Fabs\u002F2312.02179)\n- [Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?](https:\u002F\u002Farxiv.org\u002Fabs\u002F2202.12837)\n- [Surface Form Competition: Why the Highest Probability Answer Isn’t Always Right](https:\u002F\u002Farxiv.org\u002Fabs\u002F2104.08315)\n\n\u003C\u002Fdetails>\n\n### Task decomposition\n\n**Tier 1**\n\n- [Supervise Process, not Outcomes](https:\u002F\u002Fought.org\u002Fupdates\u002F2022-04-06-process)\n- [Supervising strong learners by amplifying weak experts](https:\u002F\u002Farxiv.org\u002Fabs\u002F1810.08575)\n\n**Tier 2**\n\n- [Tree of Thoughts: Deliberate Problem Solving with Large Language Models](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.10601)\n- [Factored cognition](https:\u002F\u002Fought.org\u002Fresearch\u002Ffactored-cognition)\n- [Iterated Distillation and Amplification](https:\u002F\u002Fai-alignment.com\u002Fiterated-distillation-and-amplification-157debfd1616)\n- [Recursively Summarizing Books with Human Feedback](https:\u002F\u002Farxiv.org\u002Fabs\u002F2109.10862)\n- [Solving math word problems with process-based and outcome-based feedback](https:\u002F\u002Farxiv.org\u002Fabs\u002F2211.14275)\n\n**Tier 3**\n\n- [Factored Verification: Detecting and Reducing Hallucination in Summaries of Academic Papers](https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.10627)\n- [Faithful Reasoning Using Large Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2208.14271)\n- [Iterated Decomposition: Improving Science Q&A by Supervising Reasoning Processes](https:\u002F\u002Farxiv.org\u002Fabs\u002F2301.01751)\n- [Language Model Cascades](https:\u002F\u002Farxiv.org\u002Fabs\u002F2207.10342)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>Tier 4+\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [Decontextualization: Making Sentences Stand-Alone](https:\u002F\u002Fdoi.org\u002F10.1162\u002Ftacl_a_00377)\n- [Factored Cognition Primer](https:\u002F\u002Fprimer.ought.org)\n- [Graph of Thoughts: Solving Elaborate Problems with Large Language Models](http:\u002F\u002Farxiv.org\u002Fabs\u002F2308.09687)\n- [Parsel: A Unified Natural Language Framework for Algorithmic Reasoning](http:\u002F\u002Farxiv.org\u002Fabs\u002F2212.10561)\n- [AI Chains: Transparent and Controllable Human-AI Interaction by Chaining Large Language Model Prompts](https:\u002F\u002Farxiv.org\u002Fabs\u002F2110.01691)\n- [Challenging BIG-Bench tasks and whether chain-of-thought can solve them](https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.09261)\n- [Evaluating Arguments One Step at a Time](https:\u002F\u002Fought.org\u002Fupdates\u002F2020-01-11-arguments)\n- [Least-to-Most Prompting Enables Complex Reasoning in Large Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2205.11822)\n- [Maieutic Prompting: Logically Consistent Reasoning with Recursive Explanations](https:\u002F\u002Farxiv.org\u002Fabs\u002F2205.11822)\n- [Measuring and narrowing the compositionality gap in language models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.03350)\n- [PAL: Program-aided Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2211.10435)\n- [ReAct: Synergizing Reasoning and Acting in Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.03629)\n- [Selection-Inference: Exploiting Large Language Models for Interpretable Logical Reasoning](https:\u002F\u002Farxiv.org\u002Fabs\u002F2205.10625)\n- [Show Your Work: Scratchpads for Intermediate Computation with Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2112.00114)\n- [Summ^N: A Multi-Stage Summarization Framework for Long Input Dialogues and Documents](https:\u002F\u002Farxiv.org\u002Fabs\u002F2110.10150)\n- [Thinksum: probabilistic reasoning over sets using large language models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.01293)\n\n\u003C\u002Fdetails>\n\n### Debate\n\n**Tier 2**\n\n- [AI safety via debate](https:\u002F\u002Fopenai.com\u002Fblog\u002Fdebate\u002F)\n\n**Tier 3**\n\n- ✨ [Avoiding Obfuscation with Prover-Estimator Debate](https:\u002F\u002Farxiv.org\u002Fabs\u002F2506.13609)\n- ✨ [Improving Factuality and Reasoning in Language Models through Multiagent Debate](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.14325)\n- ✨ [Prover-Verifier Games Improve Legibility of LLM Outputs](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.13692)\n- [Debate Helps Supervise Unreliable Experts](https:\u002F\u002Ftwitter.com\u002Fjoshua_clymer\u002Fstatus\u002F1724851456967417872)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>Tier 4+\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [Scalable AI Safety via Doubly-Efficient Debate](http:\u002F\u002Farxiv.org\u002Fabs\u002F2311.14125)\n- [Two-Turn Debate Doesn’t Help Humans Answer Hard Reading Comprehension Questions](https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.10860)\n\n\u003C\u002Fdetails>\n\n### Tool use and scaffolding\n\n**Tier 2**\n\n- [Measuring the impact of post-training enhancements](https:\u002F\u002Fmetr.github.io\u002Fautonomy-evals-guide\u002Felicitation-gap\u002F)\n- [WebGPT: Browser-assisted question-answering with human feedback](https:\u002F\u002Farxiv.org\u002Fabs\u002F2112.09332)\n\n**Tier 3**\n\n- ✨ [Executable Code Actions Elicit Better LLM Agents](https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.01030)\n- ✨ [GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning](https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.19457)\n- ✨ [TextGrad: Automatic \"Differentiation\" via Text](https:\u002F\u002Farxiv.org\u002Fabs\u002F2406.07496)\n- [AI capabilities can be significantly improved without expensive retraining](http:\u002F\u002Farxiv.org\u002Fabs\u002F2312.07413)\n- [Automated Statistical Model Discovery with Language Models](http:\u002F\u002Farxiv.org\u002Fabs\u002F2402.17879)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>Tier 4+\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines](http:\u002F\u002Farxiv.org\u002Fabs\u002F2310.03714)\n- [Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution](http:\u002F\u002Farxiv.org\u002Fabs\u002F2309.16797)\n- [Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation](https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.02304v1)\n- [Voyager: An Open-Ended Embodied Agent with Large Language Models](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.16291)\n- [ReGAL: Refactoring Programs to Discover Generalizable Abstractions](http:\u002F\u002Farxiv.org\u002Fabs\u002F2401.16467)\n\n\u003C\u002Fdetails>\n\n### Honesty, factuality, and epistemics\n\n**Tier 2**\n\n- [Self-critiquing models for assisting human evaluators](https:\u002F\u002Farxiv.org\u002Fabs\u002F2206.05802v2)\n\n**Tier 3**\n\n- [What Evidence Do Language Models Find Convincing?](http:\u002F\u002Farxiv.org\u002Fabs\u002F2402.11782)\n- [How to Catch an AI Liar: Lie Detection in Black-Box LLMs by Asking Unrelated Questions](https:\u002F\u002Farxiv.org\u002Fabs\u002F2309.15840)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>Tier 4+\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.04388)\n- [Long-form factuality in large language models](http:\u002F\u002Farxiv.org\u002Fabs\u002F2403.18802)\n\n\u003C\u002Fdetails>\n\n## Applications\n\n### Science\n\n**Tier 2**\n\n- ✨ [AlphaEvolve: A coding agent for scientific and algorithmic discovery](https:\u002F\u002Farxiv.org\u002Fabs\u002F2506.13131)\n- ✨ [AlphaFold 3: Accurate structure prediction of biomolecular interactions](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41586-024-07487-w)\n\n**Tier 3**\n\n- ✨ [Towards an AI Co-Scientist](https:\u002F\u002Fstorage.googleapis.com\u002Fdeepmind-media\u002FDeepMind.com\u002FBlog\u002Ftowards-an-ai-co-scientist\u002FTowards_an_AI_Co-Scientist.pdf)\n- [Can large language models provide useful feedback on research papers? A large-scale empirical analysis](http:\u002F\u002Farxiv.org\u002Fabs\u002F2310.01783)\n- [Large Language Models Encode Clinical Knowledge](http:\u002F\u002Farxiv.org\u002Fabs\u002F2212.13138)\n- [The Impact of Large Language Models on Scientific Discovery: a Preliminary Study using GPT-4](http:\u002F\u002Farxiv.org\u002Fabs\u002F2311.07361)\n- [A Dataset of Information-Seeking Questions and Answers Anchored in Research Papers](https:\u002F\u002Farxiv.org\u002Fabs\u002F2105.03011)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>Tier 4+\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [Can Generalist Foundation Models Outcompete Special-Purpose Tuning? Case Study in Medicine](http:\u002F\u002Farxiv.org\u002Fabs\u002F2311.16452)\n- [Nougat: Neural Optical Understanding for Academic Documents](http:\u002F\u002Farxiv.org\u002Fabs\u002F2308.13418)\n- [Scim: Intelligent Skimming Support for Scientific Papers](http:\u002F\u002Farxiv.org\u002Fabs\u002F2205.04561)\n- [SynerGPT: In-Context Learning for Personalized Drug Synergy Prediction and Drug Design](https:\u002F\u002Fwww.biorxiv.org\u002Fcontent\u002F10.1101\u002F2023.07.06.547759v1)\n- [Towards Accurate Differential Diagnosis with Large Language Models](http:\u002F\u002Farxiv.org\u002Fabs\u002F2312.00164)\n- [Towards a Benchmark for Scientific Understanding in Humans and Machines](http:\u002F\u002Farxiv.org\u002Fabs\u002F2304.10327)\n- [A Search Engine for Discovery of Scientific Challenges and Directions](https:\u002F\u002Farxiv.org\u002Fabs\u002F2108.13751)\n- [A full systematic review was completed in 2 weeks using automation tools: a case study](https:\u002F\u002Fpubmed.ncbi.nlm.nih.gov\u002F32004673\u002F)\n- [Fact or Fiction: Verifying Scientific Claims](https:\u002F\u002Farxiv.org\u002Fabs\u002F2004.14974)\n- [Multi-XScience: A Large-scale Dataset for Extreme Multi-document Summarization of Scientific Articles](https:\u002F\u002Farxiv.org\u002Fabs\u002F2010.14235)\n- [PEER: A Collaborative Language Model](https:\u002F\u002Farxiv.org\u002Fabs\u002F2208.11663)\n- [PubMedQA: A Dataset for Biomedical Research Question Answering](https:\u002F\u002Farxiv.org\u002Fabs\u002F1909.06146)\n- [SciCo: Hierarchical Cross-Document Coreference for Scientific Concepts](https:\u002F\u002Farxiv.org\u002Fabs\u002F2104.08809)\n- [SciTail: A Textual Entailment Dataset from Science Question Answering](http:\u002F\u002Fai2-website.s3.amazonaws.com\u002Fteam\u002Fashishs\u002Fscitail-aaai2018.pdf)\n\n\u003C\u002Fdetails>\n\n### Forecasting\n\n**Tier 3**\n\n- ✨ [Consistency Checks for Language Model Forecasters](https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.18544)\n- ✨ [LLM Processes: Numerical Predictive Distributions Conditioned on Natural Language](https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.12856)\n- [AI-Augmented Predictions: LLM Assistants Improve Human Forecasting Accuracy](https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.07862v1)\n- [Approaching Human-Level Forecasting with Language Models](http:\u002F\u002Farxiv.org\u002Fabs\u002F2402.18563)\n- [Forecasting Future World Events with Neural Networks](https:\u002F\u002Farxiv.org\u002Fabs\u002F2206.15474)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>Tier 4+\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [Are Transformers Effective for Time Series Forecasting?](https:\u002F\u002Farxiv.org\u002Fabs\u002F2205.13504)\n\n\u003C\u002Fdetails>\n\n### Search and ranking\n\n**Tier 2**\n\n- [Learning Dense Representations of Phrases at Scale](https:\u002F\u002Farxiv.org\u002Fabs\u002F2012.12624)\n- [Text and Code Embeddings by Contrastive Pre-Training](https:\u002F\u002Farxiv.org\u002Fabs\u002F2201.10005) (OpenAI embeddings)\n\n**Tier 3**\n\n- [Large Language Models are Effective Text Rankers with Pairwise Ranking Prompting](http:\u002F\u002Farxiv.org\u002Fabs\u002F2306.17563)\n- [Not All Vector Databases Are Made Equal](https:\u002F\u002Fdmitry-kan.medium.com\u002Fmilvus-pinecone-vespa-weaviate-vald-gsi-what-unites-these-buzz-words-and-what-makes-each-9c65a3bd0696)\n- [REALM: Retrieval-Augmented Language Model Pre-Training](https:\u002F\u002Farxiv.org\u002Fabs\u002F2002.08909)\n- [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https:\u002F\u002Farxiv.org\u002Fabs\u002F2005.11401)\n- [Task-aware Retrieval with Instructions](https:\u002F\u002Farxiv.org\u002Fabs\u002F2211.09260)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>Tier 4+\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [RankZephyr: Effective and Robust Zero-Shot Listwise Reranking is a Breeze!](http:\u002F\u002Farxiv.org\u002Fabs\u002F2312.02724)\n- [Some Common Mistakes In IR Evaluation, And How They Can Be Avoided](https:\u002F\u002Fdl.acm.org\u002Fdoi\u002F10.1145\u002F3190580.3190586)\n- [Boosting Search Engines with Interactive Agents](https:\u002F\u002Farxiv.org\u002Fabs\u002F2109.00527)\n- [ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT](https:\u002F\u002Farxiv.org\u002Fabs\u002F2004.12832)\n- [Moving Beyond Downstream Task Accuracy for Information Retrieval Benchmarking](https:\u002F\u002Farxiv.org\u002Fabs\u002F2212.01340)\n- [UnifiedQA: Crossing Format Boundaries With a Single QA System](https:\u002F\u002Farxiv.org\u002Fabs\u002F2005.00700)\n\n\u003C\u002Fdetails>\n\n## ML in practice\n\n### Production deployment\n\n**Tier 1**\n\n- [Machine Learning in Python: Main developments and technology trends in data science, machine learning, and AI](https:\u002F\u002Farxiv.org\u002Fabs\u002F2002.04803v2)\n- [Machine Learning: The High Interest Credit Card of Technical Debt](https:\u002F\u002Fpapers.nips.cc\u002Fpaper\u002F2015\u002Ffile\u002F86df7dcfd896fcaf2674f757a2463eba-Paper.pdf)\n\n**Tier 2**\n\n- [Designing Data-Intensive Applications](https:\u002F\u002Fdataintensive.net\u002F)\n- [A Recipe for Training Neural Networks](http:\u002F\u002Fkarpathy.github.io\u002F2019\u002F04\u002F25\u002Frecipe\u002F) (Karpathy)\n\n### Benchmarks\n\n**Tier 2**\n\n- ✨ [GAIA: a benchmark for General AI Assistants](http:\u002F\u002Farxiv.org\u002Fabs\u002F2311.12983)\n- [GPQA: A Graduate-Level Google-Proof Q&A Benchmark](http:\u002F\u002Farxiv.org\u002Fabs\u002F2311.12022)\n- [SWE-bench: Can Language Models Resolve Real-World GitHub Issues?](https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.06770v1)\n- [TruthfulQA: Measuring How Models Mimic Human Falsehoods](https:\u002F\u002Farxiv.org\u002Fabs\u002F2109.07958)\n\n**Tier 3**\n\n- ✨ [RE-Bench: Evaluating Frontier AI R&D Capabilities](https:\u002F\u002Farxiv.org\u002Fabs\u002F2411.15114)\n- ✨ [SimpleQA: Measuring Short-Form Factuality](https:\u002F\u002Fopenai.com\u002Findex\u002Fintroducing-simpleqa\u002F)\n- ✨ [ARC Prize 2024: Technical Report](https:\u002F\u002Farcprize.org\u002Fblog\u002Foai-o3-pub-breakthrough)\n- ✨ [FrontierMath: A Benchmark for Evaluating Advanced Mathematical Reasoning in AI](https:\u002F\u002Farxiv.org\u002Fabs\u002F2411.04872)\n- [Measuring Massive Multitask Language Understanding](https:\u002F\u002Farxiv.org\u002Fabs\u002F2009.03300)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>Tier 4+\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [FLEX: Unifying Evaluation for Few-Shot NLP](https:\u002F\u002Farxiv.org\u002Fabs\u002F2107.07170)\n- [Holistic Evaluation of Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2107.07170) (HELM)\n- [True Few-Shot Learning with Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2105.11447)\n- [ConditionalQA: A Complex Reading Comprehension Dataset with Conditional Answers](https:\u002F\u002Farxiv.org\u002Fabs\u002F2110.06884)\n- [Measuring Mathematical Problem Solving With the MATH Dataset](https:\u002F\u002Farxiv.org\u002Fabs\u002F2103.03874)\n- [QuALITY: Question Answering with Long Input Texts, Yes!](https:\u002F\u002Farxiv.org\u002Fabs\u002F2112.08608)\n- [SCROLLS: Standardized CompaRison Over Long Language Sequences](https:\u002F\u002Farxiv.org\u002Fabs\u002F2201.03533)\n- [What Will it Take to Fix Benchmarking in Natural Language Understanding?](https:\u002F\u002Farxiv.org\u002Fabs\u002F2104.02145)\n\n\u003C\u002Fdetails>\n\n### Datasets\n\n**Tier 2**\n\n- [Common Crawl](https:\u002F\u002Farxiv.org\u002Fabs\u002F2105.02732)\n- [The Pile: An 800GB Dataset of Diverse Text for Language Modeling](https:\u002F\u002Farxiv.org\u002Fabs\u002F2101.00027)\n\n**Tier 3**\n\n- ✨ [FineWeb: Decanting the Web for the Finest Text Data at Scale](https:\u002F\u002Fhuggingface.co\u002Fspaces\u002FHuggingFaceFW\u002Fblogpost-fineweb-v1)\n- [Dialog Inpainting: Turning Documents into Dialogs](https:\u002F\u002Farxiv.org\u002Fabs\u002F2205.09073)\n- [MS MARCO: A Human Generated MAchine Reading COmprehension Dataset](https:\u002F\u002Farxiv.org\u002Fabs\u002F1611.09268)\n- [Microsoft Academic Graph](https:\u002F\u002Finternal-journal.frontiersin.org\u002Farticles\u002F10.3389\u002Ffdata.2019.00045\u002Ffull)\n\n## Advanced topics\n\n### World models and causality\n\n**Tier 3**\n\n- [Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task](http:\u002F\u002Farxiv.org\u002Fabs\u002F2210.13382)\n- [From Word Models to World Models: Translating from Natural Language to the Probabilistic Language of Thought](http:\u002F\u002Farxiv.org\u002Fabs\u002F2306.12672)\n- [Language Models Represent Space and Time](http:\u002F\u002Farxiv.org\u002Fabs\u002F2310.02207)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>Tier 4+\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [Amortizing intractable inference in large language models](http:\u002F\u002Farxiv.org\u002Fabs\u002F2310.04363)\n- [CLADDER: Assessing Causal Reasoning in Language Models](http:\u002F\u002Fzhijing-jin.com\u002Ffiles\u002Fpapers\u002FCLadder_2023.pdf)\n- [Causal Bayesian Optimization](https:\u002F\u002Fproceedings.mlr.press\u002Fv108\u002Faglietti20a.html)\n- [Causal Reasoning and Large Language Models: Opening a New Frontier for Causality](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.00050)\n- [Generative Agents: Interactive Simulacra of Human Behavior](http:\u002F\u002Farxiv.org\u002Fabs\u002F2304.03442)\n- [Passive learning of active causal strategies in agents and language models](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.16183)\n\n\u003C\u002Fdetails>\n\n### Planning\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>Tier 4+\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [Beyond A\\*: Better Planning with Transformers via Search Dynamics Bootstrapping](http:\u002F\u002Farxiv.org\u002Fabs\u002F2402.14083)\n- [Cognitive Architectures for Language Agents](http:\u002F\u002Farxiv.org\u002Fabs\u002F2309.02427)\n\n\u003C\u002Fdetails>\n\n### Uncertainty, calibration, and active learning\n\n**Tier 2**\n\n- [Experts Don't Cheat: Learning What You Don't Know By Predicting Pairs](http:\u002F\u002Farxiv.org\u002Fabs\u002F2402.08733)\n- [A Simple Baseline for Bayesian Uncertainty in Deep Learning](https:\u002F\u002Farxiv.org\u002Fabs\u002F1902.02476)\n- [Plex: Towards Reliability using Pretrained Large Model Extensions](https:\u002F\u002Farxiv.org\u002Fabs\u002F2207.07411)\n\n**Tier 3**\n\n- ✨ [Textual Bayes: Quantifying Uncertainty in LLM-Based Systems](https:\u002F\u002Farxiv.org\u002Fabs\u002F2506.10060)\n- [Active Preference Inference using Language Models and Probabilistic Reasoning](http:\u002F\u002Farxiv.org\u002Fabs\u002F2312.12009)\n- [Eliciting Human Preferences with Language Models](http:\u002F\u002Farxiv.org\u002Fabs\u002F2310.11589)\n- [Describing Differences between Text Distributions with Natural Language](https:\u002F\u002Farxiv.org\u002Fabs\u002F2201.12323)\n- [Teaching Models to Express Their Uncertainty in Words](https:\u002F\u002Farxiv.org\u002Fabs\u002F2205.14334)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>Tier 4+\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [Active Learning by Acquiring Contrastive Examples](https:\u002F\u002Farxiv.org\u002Fabs\u002F2109.03764)\n- [Doing Experiments and Revising Rules with Natural Language and Probabilistic Reasoning](http:\u002F\u002Farxiv.org\u002Fabs\u002F2402.06025)\n- [STaR-GATE: Teaching Language Models to Ask Clarifying Questions](http:\u002F\u002Farxiv.org\u002Fabs\u002F2403.19154)\n- [Active Testing: Sample-Efficient Model Evaluation](https:\u002F\u002Farxiv.org\u002Fabs\u002F2103.05331)\n- [Uncertainty Estimation for Language Reward Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2203.07472)\n\n\u003C\u002Fdetails>\n\n### Interpretability and model editing\n\n**Tier 2**\n\n- ✨ [Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet](https:\u002F\u002Ftransformer-circuits.pub\u002F2024\u002Fscaling-monosemanticity\u002F)\n- ✨ [Interpretability at Scale: Identifying Causal Mechanisms in Alpaca](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.08809)\n- [Discovering Latent Knowledge in Language Models Without Supervision](https:\u002F\u002Farxiv.org\u002Fabs\u002F2212.03827v1)\n\n**Tier 3**\n\n- ✨ [Scaling and Evaluating Sparse Autoencoders](https:\u002F\u002Fcdn.openai.com\u002Fpapers\u002Fsparse-autoencoders.pdf)\n- ✨ [Opening the AI black box: program synthesis via mechanistic interpretability](https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.05110v1)\n- [Mechanistically analyzing the effects of fine-tuning on procedurally defined tasks](http:\u002F\u002Farxiv.org\u002Fabs\u002F2311.12786)\n- [Representation Engineering: A Top-Down Approach to AI Transparency](http:\u002F\u002Farxiv.org\u002Fabs\u002F2310.01405)\n- [Studying Large Language Model Generalization with Influence Functions](http:\u002F\u002Farxiv.org\u002Fabs\u002F2308.03296)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>Tier 4+\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [Codebook Features: Sparse and Discrete Interpretability for Neural Networks](http:\u002F\u002Farxiv.org\u002Fabs\u002F2310.17230)\n- [Eliciting Latent Predictions from Transformers with the Tuned Lens](http:\u002F\u002Farxiv.org\u002Fabs\u002F2303.08112)\n- [How do Language Models Bind Entities in Context?](http:\u002F\u002Farxiv.org\u002Fabs\u002F2310.17191)\n- [Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small](https:\u002F\u002Farxiv.org\u002Fabs\u002F2211.00593)\n- [Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models](http:\u002F\u002Farxiv.org\u002Fabs\u002F2403.19647)\n- [Uncovering mesa-optimization algorithms in Transformers](http:\u002F\u002Farxiv.org\u002Fabs\u002F2309.05858)\n- [Fast Model Editing at Scale](https:\u002F\u002Farxiv.org\u002Fabs\u002F2110.11309)\n- [Git Re-Basin: Merging Models modulo Permutation Symmetries](https:\u002F\u002Farxiv.org\u002Fabs\u002F2209.04836)\n- [Locating and Editing Factual Associations in GPT](https:\u002F\u002Farxiv.org\u002Fabs\u002F2202.05262)\n- [Mass-Editing Memory in a Transformer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.07229)\n\n\u003C\u002Fdetails>\n\n### Reinforcement learning\n\n**Tier 2**\n\n- ✨ [DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.03300) (GRPO)\n- [Direct Preference Optimization: Your Language Model is Secretly a Reward Model](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.18290)\n- [Reflexion: Language Agents with Verbal Reinforcement Learning](http:\u002F\u002Farxiv.org\u002Fabs\u002F2303.11366)\n- [Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm](https:\u002F\u002Farxiv.org\u002Fabs\u002F1712.01815) (AlphaZero)\n- [MuZero: Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model](https:\u002F\u002Farxiv.org\u002Fabs\u002F1911.08265)\n\n**Tier 3**\n\n- [Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback](http:\u002F\u002Farxiv.org\u002Fabs\u002F2307.15217)\n- [AlphaStar: mastering the real-time strategy game StarCraft II](https:\u002F\u002Fdeepmind.com\u002Fblog\u002Farticle\u002Falphastar-mastering-real-time-strategy-game-starcraft-ii)\n- [Decision Transformer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2106.01345)\n- [Mastering Atari Games with Limited Data](https:\u002F\u002Farxiv.org\u002Fabs\u002F2111.00210) (EfficientZero)\n- [Mastering Stratego, the classic game of imperfect information](https:\u002F\u002Fwww.science.org\u002Fdoi\u002F10.1126\u002Fscience.add4679) (DeepNash)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>Tier 4+\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [AlphaStar Unplugged: Large-Scale Offline Reinforcement Learning](http:\u002F\u002Farxiv.org\u002Fabs\u002F2308.03526)\n- [Bayesian Reinforcement Learning with Limited Cognitive Load](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.03263)\n- [Contrastive Prefence Learning: Learning from Human Feedback without RL](http:\u002F\u002Farxiv.org\u002Fabs\u002F2310.13639)\n- [Grandmaster-Level Chess Without Search](http:\u002F\u002Farxiv.org\u002Fabs\u002F2402.04494)\n- [A data-driven approach for learning to control computers](https:\u002F\u002Farxiv.org\u002Fabs\u002F2202.08137)\n- [Acquisition of Chess Knowledge in AlphaZero](https:\u002F\u002Farxiv.org\u002Fabs\u002F2111.09259)\n- [Player of Games](https:\u002F\u002Farxiv.org\u002Fabs\u002F2112.03178)\n- [Retrieval-Augmented Reinforcement Learning](https:\u002F\u002Farxiv.org\u002Fabs\u002F2202.08417)\n\n\u003C\u002Fdetails>\n\n## The big picture\n\n### AI scaling\n\n**Tier 1**\n\n- [Scaling Laws for Neural Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2001.08361)\n- [Takeoff speeds](https:\u002F\u002Fsideways-view.com\u002F2018\u002F02\u002F24\u002Ftakeoff-speeds\u002F)\n- [The Bitter Lesson](http:\u002F\u002Fwww.incompleteideas.net\u002FIncIdeas\u002FBitterLesson.html)\n\n**Tier 2**\n\n- [AI and compute](https:\u002F\u002Fopenai.com\u002Fblog\u002Fai-and-compute\u002F)\n- [Scaling Laws for Transfer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2102.01293)\n- [Training Compute-Optimal Large Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2203.15556) (Chinchilla)\n\n**Tier 3**\n\n- ✨ [Pre-training under Infinite Compute](https:\u002F\u002Farxiv.org\u002Fabs\u002F2509.14786)\n- [Emergent Abilities of Large Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2206.07682)\n- [Transcending Scaling Laws with 0.1% Extra Compute](https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.11399) (U-PaLM)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>Tier 4+\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [Physics of Language Models: Part 3.3, Knowledge Capacity Scaling Laws](http:\u002F\u002Farxiv.org\u002Fabs\u002F2404.05405)\n- [Power Law Trends in Speedrunning and Machine Learning](http:\u002F\u002Farxiv.org\u002Fabs\u002F2304.10004)\n- [Scaling laws for single-agent reinforcement learning](http:\u002F\u002Farxiv.org\u002Fabs\u002F2301.13442)\n- [Beyond neural scaling laws: beating power law scaling via data pruning](https:\u002F\u002Farxiv.org\u002Fabs\u002F2206.14486)\n- [Emergent Abilities of Large Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2206.07682)\n- [Scaling Scaling Laws with Board Games](https:\u002F\u002Farxiv.org\u002Fabs\u002F2104.03113)\n\n\u003C\u002Fdetails>\n\n### AI safety\n\n**Tier 1**\n\n- [Three impacts of machine intelligence](https:\u002F\u002Fwww.effectivealtruism.org\u002Farticles\u002Fthree-impacts-of-machine-intelligence-paul-christiano\u002F)\n- [What failure looks like](https:\u002F\u002Fwww.alignmentforum.org\u002Fposts\u002FHBxe6wdjxK239zajf\u002Fwhat-failure-looks-like)\n- [Without specific countermeasures, the easiest path to transformative AI likely leads to AI takeover](https:\u002F\u002Fwww.lesswrong.com\u002Fposts\u002FpRkFkzwKZ2zfa3R6H\u002Fwithout-specific-countermeasures-the-easiest-path-to)\n\n**Tier 2**\n\n- [An Overview of Catastrophic AI Risks](http:\u002F\u002Farxiv.org\u002Fabs\u002F2306.12001)\n- [Clarifying “What failure looks like” (part 1)](https:\u002F\u002Fwww.lesswrong.com\u002Fposts\u002Fv6Q7T335KCMxujhZu\u002Fclarifying-what-failure-looks-like-part-1)\n- [Deep RL from human preferences](https:\u002F\u002Fopenai.com\u002Fblog\u002Fdeep-reinforcement-learning-from-human-preferences\u002F)\n- [The alignment problem from a deep learning perspective](https:\u002F\u002Farxiv.org\u002Fabs\u002F2209.00626)\n\n**Tier 3**\n\n- ✨ [Alignment Faking in Large Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.14093)\n- ✨ [Constitutional Classifiers: Defending against Universal Jailbreaks](https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.18837)\n- ✨ [Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs](https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.17424)\n- ✨ [Gradual Disempowerment: Systemic Existential Risks from Incremental AI Development](https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.16946)\n- [Scheming AIs: Will AIs fake alignment during training in order to get power?](http:\u002F\u002Farxiv.org\u002Fabs\u002F2311.08379)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>Tier 4+\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [Towards a Law of Iterated Expectations for Heuristic Estimators](https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.01290)\n- [Measuring Progress on Scalable Oversight for Large Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2211.03540)\n- [Scalable agent alignment via reward modelling](https:\u002F\u002Farxiv.org\u002Fabs\u002F1811.07871)\n- [AI Deception: A Survey of Examples, Risks, and Potential Solutions](http:\u002F\u002Farxiv.org\u002Fabs\u002F2308.14752)\n- [Benchmarks for Detecting Measurement Tampering](http:\u002F\u002Farxiv.org\u002Fabs\u002F2308.15605)\n- [Chess as a Testing Grounds for the Oracle Approach to AI Safety](http:\u002F\u002Farxiv.org\u002Fabs\u002F2010.02911)\n- [Close the Gates to an Inhuman Future: How and why we should choose to not develop superhuman general-purpose artificial intelligence](https:\u002F\u002Fpapers.ssrn.com\u002Fabstract=4608505)\n- [Model evaluation for extreme risks](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.15324)\n- [Responsible Reporting for Frontier AI Development](http:\u002F\u002Farxiv.org\u002Fabs\u002F2404.02675)\n- [Safety Cases: How to Justify the Safety of Advanced AI Systems](http:\u002F\u002Farxiv.org\u002Fabs\u002F2403.10462)\n- [Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training](http:\u002F\u002Farxiv.org\u002Fabs\u002F2401.05566)\n- [Technical Report: Large Language Models can Strategically Deceive their Users when Put Under Pressure](http:\u002F\u002Farxiv.org\u002Fabs\u002F2311.07590)\n- [Tensor Trust: Interpretable Prompt Injection Attacks from an Online Game](http:\u002F\u002Farxiv.org\u002Fabs\u002F2311.01011)\n- [Tools for Verifying Neural Models' Training Data](http:\u002F\u002Farxiv.org\u002Fabs\u002F2307.00682)\n- [Towards a Cautious Scientist AI with Convergent Safety Bounds](https:\u002F\u002Fyoshuabengio.org\u002F2024\u002F02\u002F26\u002Ftowards-a-cautious-scientist-ai-with-convergent-safety-bounds\u002F)\n- [Alignment of Language Agents](https:\u002F\u002Farxiv.org\u002Fabs\u002F2103.14659)\n- [Eliciting Latent Knowledge](https:\u002F\u002Fdocs.google.com\u002Fdocument\u002Fd\u002F1WwsnJQstPq91_Yh-Ch2XRL8H_EpsnjrC1dwZXR37PC8\u002Fedit?usp=sharing)\n- [Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned](https:\u002F\u002Farxiv.org\u002Fabs\u002F2209.07858)\n- [Red Teaming Language Models with Language Models](https:\u002F\u002Fstorage.googleapis.com\u002Fdeepmind-media\u002FRed%20Teaming\u002FRed%20Teaming.pdf)\n- [Risks from Learned Optimization in Advanced Machine Learning Systems](https:\u002F\u002Farxiv.org\u002Fabs\u002F1906.01820)\n- [Unsolved Problems in ML Safety](https:\u002F\u002Farxiv.org\u002Fabs\u002F2109.13916)\n\n\u003C\u002Fdetails>\n\n### Economic and social impacts\n\n**Tier 2**\n\n- ✨ [AI 2027](https:\u002F\u002Fai-2027.com\u002F)\n- ✨ [Situational Awareness](https:\u002F\u002Fsituational-awareness.ai\u002F) (Aschenbrenner)\n\n**Tier 3**\n\n- [Explosive growth from AI automation: A review of the arguments](http:\u002F\u002Farxiv.org\u002Fabs\u002F2309.11690)\n- [Language Models Can Reduce Asymmetry in Information Markets](http:\u002F\u002Farxiv.org\u002Fabs\u002F2403.14443)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>Tier 4+\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [Bridging the Human-AI Knowledge Gap: Concept Discovery and Transfer in AlphaZero](http:\u002F\u002Farxiv.org\u002Fabs\u002F2310.16410)\n- [Foundation Models and Fair Use](https:\u002F\u002Farxiv.org\u002Fabs\u002F2303.15715v1)\n- [GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models](http:\u002F\u002Farxiv.org\u002Fabs\u002F2303.10130)\n- [Levels of AGI: Operationalizing Progress on the Path to AGI](http:\u002F\u002Farxiv.org\u002Fabs\u002F2311.02462)\n- [Opportunities and Risks of LLMs for Scalable Deliberation with Polis](http:\u002F\u002Farxiv.org\u002Fabs\u002F2306.11932)\n- [On the Opportunities and Risks of Foundation Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2108.07258)\n\n\u003C\u002Fdetails>\n\n### Philosophy\n\n**Tier 2**\n\n- [Meaning without reference in large language models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2208.02957)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>Tier 4+\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [Consciousness in Artificial Intelligence: Insights from the Science of Consciousness](http:\u002F\u002Farxiv.org\u002Fabs\u002F2308.08708)\n- [Philosophers Ought to Develop, Theorize About, and Use Philosophically Relevant AI](https:\u002F\u002Fphilarchive.org\u002Farchive\u002FCLAPOT-16)\n- [Towards Evaluating AI Systems for Moral Status Using Self-Reports](http:\u002F\u002Farxiv.org\u002Fabs\u002F2311.08576)\n\n\u003C\u002Fdetails>\n\n## Maintainer\n\n[andreas@elicit.com](mailto:andreas@elicit.com)\n","# Elicit 机器学习阅读清单\n\n## 目的\n\n本课程旨在帮助 Elicit 的新员工学习机器学习的基础知识，重点关注语言模型。我试图在适用于生产环境中部署机器学习的相关论文与对长期可扩展性至关重要的技术之间取得平衡。\n\n如果你尚未加入 Elicit——我们正在招聘机器学习工程师和软件工程师（[点击此处查看职位](https:\u002F\u002Felicit.com\u002Fcareers)）。\n\n## 阅读指南\n\n推荐阅读顺序：\n\n1. 所有主题先读“Tier 1”\n2. 再读“Tier 2”\n3. 以此类推\n\n✨ = 2025年11月26日之后新增\n\n## 目录\n\n- [基础概念](#fundamentals)\n  - [机器学习导论](#introduction-to-machine-learning)\n  - [Transformer 模型](#transformers)\n  - [关键的基础模型架构](#key-foundation-model-architectures)\n  - [训练与微调](#training-and-finetuning)\n- [推理与运行时策略](#reasoning-and-runtime-strategies)\n  - [上下文推理](#in-context-reasoning)\n  - [任务分解](#task-decomposition)\n  - [辩论](#debate)\n  - [工具使用与脚手架](#tool-use-and-scaffolding)\n  - [诚实性、事实性和认识论](#honesty-factuality-and-epistemics)\n- [应用](#applications)\n  - [科学](#science)\n  - [预测](#forecasting)\n  - [搜索与排序](#search-and-ranking)\n- [机器学习实战](#ml-in-practice)\n  - [生产环境部署](#production-deployment)\n  - [基准测试](#benchmarks)\n  - [数据集](#datasets)\n- [高级主题](#advanced-topics)\n  - [世界模型与因果关系](#world-models-and-causality)\n  - [规划](#planning)\n  - [不确定性、校准与主动学习](#uncertainty-calibration-and-active-learning)\n  - [可解释性与模型编辑](#interpretability-and-model-editing)\n  - [强化学习](#reinforcement-learning)\n- [宏观视角](#the-big-picture)\n  - [AI 扩展规律](#ai-scaling)\n  - [AI 安全](#ai-safety)\n  - [经济与社会影响](#economic-and-social-impacts)\n  - [哲学](#philosophy)\n- [维护者](#maintainer)\n\n## 基础概念\n\n### 机器学习导论\n\n**Tier 1**\n\n- [机器学习简短介绍](https:\u002F\u002Fwww.alignmentforum.org\u002Fposts\u002FqE73pqxAZmeACsAdF\u002Fa-short-introduction-to-machine-learning)\n- [那么，什么是神经网络？](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=IHZwWFHWa-w)\n- [梯度下降：神经网络的学习机制](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=Ilg3gGewQ5U&list=PLZHQObOWTQDNU6R1_67000Dx_ZCJB-3pi&index=4)\n\n**Tier 2**\n\n- [反向传播的直观理解](https:\u002F\u002Fcs231n.github.io\u002Foptimization-2\u002F)\n- [反向传播到底在做什么？](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=tIeHLnjs5U8&list=PLZHQObOWTQDNU6R1_67000Dx_ZCJB-3pi&index=5)\n- [深度强化学习入门](https:\u002F\u002Fthomassimonini.medium.com\u002Fan-introduction-to-deep-reinforcement-learning-17a565999c0c)\n\n**Tier 3**\n\n- [神经网络与反向传播详解：从零构建 micrograd](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=VMj-3S1tku0)（Karpathy）\n- [反向传播的数学推导](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=eMlx5fFNoYc&list=PLZHQObOWTQDNU6R1_67000Dx_ZCJB-3pi&index=5)\n\n### Transformer 模型\n\n**Tier 1**\n\n- ✨ [大型语言模型简介](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=zjkBMFhNj_g)（Karpathy）\n- [那么，什么是 GPT？Transformer 的可视化入门](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=wjZofJX0v4M)\n- [Transformer 中的注意力机制：可视化讲解](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=eMlx5fFNoYc)\n- [注意力？注意！](https:\u002F\u002Flilianweng.github.io\u002Fposts\u002F2018-06-24-attention\u002F)\n- [图解 Transformer](http:\u002F\u002Fjalammar.github.io\u002Fillustrated-transformer\u002F)\n\n**Tier 2**\n\n- ✨ [深入解析 ChatGPT 等大型语言模型](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=7xTGNNLPyMI)（Karpathy）\n- [让我们一起构建 GPT 分词器](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=zduSFxRajkE)（Karpathy）\n- [图解 GPT-2：可视化 Transformer 语言模型](https:\u002F\u002Fjalammar.github.io\u002Fillustrated-gpt2\u002F)\n- [通过联合学习对齐与翻译实现神经机器翻译](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1409.0473)\n- [Attention Is All You Need](https:\u002F\u002Farxiv.org\u002Fabs\u002F1706.03762)\n\n**Tier 3**\n\n- ✨ [反转诅咒：以“A 是 B”训练的语言模型无法学会“B 是 A”](https:\u002F\u002Farxiv.org\u002Fabs\u002F2309.12288)\n- [注释版 Transformer](https:\u002F\u002Fnlp.seas.harvard.edu\u002F2018\u002F04\u002F03\u002Fattention.html)\n- [TabPFN：能在一秒内解决小型表格分类问题的 Transformer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2207.01848)\n- [Grokking：超越过拟合的小型算法数据集上的泛化能力](https:\u002F\u002Farxiv.org\u002Fabs\u002F2201.02177)\n- [Transformer 电路的数学框架](https:\u002F\u002Ftransformer-circuits.pub\u002F2021\u002Fframework\u002Findex.html)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>Tier 4 及以上\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [关于更快更轻量级 Transformer 的实用综述](https:\u002F\u002Farxiv.org\u002Fabs\u002F2103.14636)\n- [自回归 Transformer 的组合能力：基于合成可解释任务的研究](https:\u002F\u002Farxiv.org\u002Fabs\u002F2311.12997)\n- [记忆型 Transformer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2203.08913)\n- [Transformer 的前馈层是键值存储](https:\u002F\u002Farxiv.org\u002Fabs\u002F2012.14913)\n\n\u003C\u002Fdetails>\n\n### 重要基础模型架构\n\n**第一层**\n\n- [语言模型是无监督的多任务学习者](https:\u002F\u002Fwww.semanticscholar.org\u002Fpaper\u002FLanguage-Models-are-Unsupervised-Multitask-Learners-Radford-Wu\u002F9405cc0d6169988371b2755e573cc28650d14dfe)（GPT-2）\n- [语言模型是少样本学习者](https:\u002F\u002Farxiv.org\u002Fabs\u002F2005.14165)（GPT-3）\n\n**第二层**\n\n- ✨ [DeepSeek-R1](https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.12948)（DeepSeek-R1）\n- ✨ [DeepSeek-V3 技术报告](https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.19437)（DeepSeek-V3）\n- ✨ [Llama 3 模型家族](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.21783)（Llama 3）\n- [LLaMA：开放且高效的基座语言模型](http:\u002F\u002Farxiv.org\u002Fabs\u002F2302.13971)（LLaMA）\n- [通过人类反馈训练语言模型遵循指令](https:\u002F\u002Farxiv.org\u002Fabs\u002F2203.02155)（OpenAI Instruct）\n\n**第三层**\n\n- ✨ [LLaMA 2：开放的基座模型与微调后的聊天模型](https:\u002F\u002Farxiv.org\u002Fabs\u002F2307.09288)（LLaMA 2）\n- ✨ [Qwen2.5 技术报告](https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.15115)（Qwen2.5）\n- ✨ [Titans：在推理时学习记忆](https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.00663)\n- ✨ [字节潜伏变换器](https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.09871)\n- ✨ [Phi-4 技术报告](https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.08905)（phi-4）\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>第四层及以上\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [评估基于代码训练的大规模语言模型](https:\u002F\u002Farxiv.org\u002Fabs\u002F2107.03374)（OpenAI Codex）\n- [Mistral 7B](http:\u002F\u002Farxiv.org\u002Fabs\u002F2310.06825)（Mistral）\n- [探索统一文本到文本变换器的迁移学习极限](https:\u002F\u002Farxiv.org\u002Fabs\u002F1910.10683)（T5）\n- [Gemini：高度强大的多模态模型家族](https:\u002F\u002Fstorage.googleapis.com\u002Fdeepmind-media\u002Fgemini\u002Fgemini_1_report.pdf)（Gemini）\n- [Mamba：具有选择性状态空间的线性时间序列建模](https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.00752v1)（Mamba）\n- [扩展指令微调语言模型](https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.11416)（Flan）\n- [利用结构化状态空间高效建模长序列](https:\u002F\u002Farxiv.org\u002Fabs\u002F2111.00396)（[视频](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=EvQ3ncuriCM)）（S4）\n- [一致性模型](http:\u002F\u002Farxiv.org\u002Fabs\u002F2303.01469)\n- [Claude 模型的模型卡片与评估](https:\u002F\u002Fwww-cdn.anthropic.com\u002Fbd2a28d2535bfb0494cc8e2a3bf135d2e7523226\u002FModel-Card-Claude-2.pdf)（Claude 2）\n- [OLMo：加速语言模型科学研究](http:\u002F\u002Farxiv.org\u002Fabs\u002F2402.00838)\n- [PaLM 2 技术报告](https:\u002F\u002Farxiv.org\u002Fabs\u002F2305.10403)（Palm 2）\n- [教科书就够了 II：phi-1.5 技术报告](http:\u002F\u002Farxiv.org\u002Fabs\u002F2309.05463)（phi 1.5）\n- [视觉指令微调](http:\u002F\u002Farxiv.org\u002Fabs\u002F2304.08485)（LLaVA）\n- [通用语言助手作为对齐的实验室](https:\u002F\u002Farxiv.org\u002Fabs\u002F2112.00861)\n- [微调后的语言模型是零样本学习者](https:\u002F\u002Farxiv.org\u002Fabs\u002F2109.01652)（Google Instruct）\n- [Galactica：面向科学的大规模语言模型](https:\u002F\u002Farxiv.org\u002Fabs\u002F2211.09085)\n- [LaMDA：用于对话应用的语言模型](https:\u002F\u002Farxiv.org\u002Fabs\u002F2201.08239)（Google Dialog）\n- [OPT：开放的预训练变换器语言模型](https:\u002F\u002Farxiv.org\u002Fabs\u002F2112.11446)（Meta GPT-3）\n- [PaLM：通过 Pathways 扩展语言建模能力](https:\u002F\u002Farxiv.org\u002Fabs\u002F2204.02311)（PaLM）\n- [利用大规模语言模型进行程序合成](https:\u002F\u002Farxiv.org\u002Fabs\u002F2108.07732)（Google Codex）\n- [扩展语言模型：方法、分析及训练 Gopher 的洞见](https:\u002F\u002Farxiv.org\u002Fabs\u002F2112.11446)（Gopher）\n- [利用语言模型解决定量推理问题](https:\u002F\u002Farxiv.org\u002Fabs\u002F2206.14858)（Minerva）\n- [UL2：统一语言学习范式](http:\u002F\u002Faima.cs.berkeley.edu\u002F)（UL2）\n\n\u003C\u002Fdetails>\n\n### 训练与微调\n\n**第二层级**\n\n- ✨ [深度学习调优手册](https:\u002F\u002Fgithub.com\u002Fgoogle-research\u002Ftuning_playbook)\n- [张量程序 V：通过零样本超参数迁移调优大型神经网络](https:\u002F\u002Farxiv.org\u002Fabs\u002F2203.03466)\n- [基于人类反馈的学习摘要生成](https:\u002F\u002Farxiv.org\u002Fabs\u002F2009.01325)\n- [训练验证器解决数学应用题](https:\u002F\u002Farxiv.org\u002Fabs\u002F2110.14168)\n\n**第三层级**\n\n- ✨ [通过多标记预测实现更好更快的大型语言模型](https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.19737)\n- ✨ [LoRA 与全量微调：等效性的幻象](https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.21228)\n- ✨ [QLoRA：量化 LLM 的高效微调](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.14314)\n- [利用人类偏好预训练语言模型](http:\u002F\u002Farxiv.org\u002Fabs\u002F2302.08582)\n- [弱监督到强能力：用弱监督激发强大能力](http:\u002F\u002Farxiv.org\u002Fabs\u002F2312.09390)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>第四层级及以上\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [少样本参数高效微调比上下文学习更好更便宜](https:\u002F\u002Farxiv.org\u002Fabs\u002F2205.05638v1)\n- [超越人类数据：扩展语言模型的问题解决自训练](http:\u002F\u002Farxiv.org\u002Fabs\u002F2312.06585)\n- [通过自然语言反馈训练提升代码生成能力](http:\u002F\u002Farxiv.org\u002Fabs\u002F2303.16749)\n- [语言建模即压缩](https:\u002F\u002Farxiv.org\u002Fabs\u002F2309.10668v1)\n- [LIMA：对齐之道，少即是多](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.11206)\n- [使用梗概标记学习压缩提示](http:\u002F\u002Farxiv.org\u002Fabs\u002F2304.08467)\n- [迷失于中间：语言模型如何利用长上下文](http:\u002F\u002Farxiv.org\u002Fabs\u002F2307.03172)\n- [LoRA：大型语言模型的低秩适应](https:\u002F\u002Farxiv.org\u002Fabs\u002F2106.09685)\n- [Quiet-STaR：语言模型可以自我训练，在开口前先思考](http:\u002F\u002Farxiv.org\u002Fabs\u002F2403.09629)\n- [用于语言建模的强化自训练（ReST）](http:\u002F\u002Farxiv.org\u002Fabs\u002F2308.08998)\n- [无需人类示范即可解决奥林匹克几何问题](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41586-023-06747-5)\n- [告知而非展示：声明性事实影响 LLM 的泛化能力](http:\u002F\u002Farxiv.org\u002Fabs\u002F2312.07779)\n- [只需教科书就够了](http:\u002F\u002Farxiv.org\u002Fabs\u002F2306.11644)\n- [TinyStories：语言模型能有多小，仍能说出连贯的英语？](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.07759)\n- [大规模使用语言反馈训练语言模型](http:\u002F\u002Farxiv.org\u002Fabs\u002F2303.16755)\n- [图灵完备的 Transformer：两个 Transformer 比一个更强大](https:\u002F\u002Fopenreview.net\u002Fforum?id=MGWsPGogLH)\n- [ByT5：迈向无标记未来——预训练字节到字节模型](https:\u002F\u002Farxiv.org\u002Fabs\u002F2105.13626)\n- [数据分布特性驱动 Transformer 中涌现的上下文学习](https:\u002F\u002Farxiv.org\u002Fabs\u002F2205.05055)\n- [Diffusion-LM 改善可控文本生成](https:\u002F\u002Farxiv.org\u002Fabs\u002F2205.14217)\n- [ERNIE 3.0：面向语言理解和生成的大规模知识增强预训练](https:\u002F\u002Farxiv.org\u002Fabs\u002F2107.02137)\n- [高效训练语言模型以填补中间空白](https:\u002F\u002Farxiv.org\u002Fabs\u002F2207.14255)\n- [ExT5：迈向极端多任务规模的迁移学习](https:\u002F\u002Farxiv.org\u002Fabs\u002F2111.10952)\n- [前缀调优：优化连续提示以进行生成](https:\u002F\u002Farxiv.org\u002Fabs\u002F2101.00190)\n- [数据点之间的自注意力：超越深度学习中的单个输入输出对](https:\u002F\u002Farxiv.org\u002Fabs\u002F2106.02584)\n- [真正的少样本学习与提示——现实视角](https:\u002F\u002Farxiv.org\u002Fabs\u002F2111.13440)\n\n\u003C\u002Fdetails>\n\n## 推理与运行时策略\n\n### 上下文推理\n\n**第二层级**\n\n- ✨ [在测试时以最佳方式扩展 LLM 的计算资源，可能比扩大模型参数更有效](https:\u002F\u002Farxiv.org\u002Fabs\u002F2408.03314)\n- [思维链提示激发大型语言模型的推理能力](https:\u002F\u002Farxiv.org\u002Fabs\u002F2201.11903)\n- [大型语言模型是零样本推理者](https:\u002F\u002Farxiv.org\u002Fabs\u002F2205.11916)（让我们一步步思考）\n- [自我一致性改进语言模型中的思维链推理](https:\u002F\u002Farxiv.org\u002Fabs\u002F2203.11171)\n\n**第三层级**\n\n- ✨ [s1：简单的测试时缩放](https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.19393)\n- ✨ [使自我改进型推理者成为可能的认知行为，或，高效率 STaR 的四种习惯](https:\u002F\u002Farxiv.org\u002Fabs\u002F2503.01307)\n- ✨ [测试时训练在抽象推理中的惊人效果](https:\u002F\u002Farxiv.org\u002Fabs\u002F2411.07279)\n- ✨ [大型语言模型目前尚无法自我纠正推理](https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.01798v1)\n- [无需提示的思维链推理](http:\u002F\u002Farxiv.org\u002Fabs\u002F2402.10200)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>第四层级及以上\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [为何要一步步思考？推理源于经验的局部性](http:\u002F\u002Farxiv.org\u002Fabs\u002F2304.03843)\n- [Baldur：利用大型语言模型进行完整证明的生成与修复](https:\u002F\u002Farxiv.org\u002Fabs\u002F2303.04910v1)\n- [偏见增强的一致性训练可减少思维链中的偏见推理](http:\u002F\u002Farxiv.org\u002Fabs\u002F2403.05518)\n- [使用语言模型进行认证推理](http:\u002F\u002Farxiv.org\u002Fabs\u002F2306.04031)\n- [假设搜索：利用语言模型进行归纳推理](http:\u002F\u002Farxiv.org\u002Fabs\u002F2309.05660)\n- [LLMs 与抽象与推理语料库：成功、失败以及基于对象表示的重要性](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.18354)\n- [搜索流（SoS）：学习在语言中进行搜索](http:\u002F\u002Farxiv.org\u002Fabs\u002F2404.03683)\n- [通过潜在变量推断训练思维链](http:\u002F\u002Farxiv.org\u002Fabs\u002F2312.02179)\n- [重新思考演示的作用：是什么让上下文学习奏效？](https:\u002F\u002Farxiv.org\u002Fabs\u002F2202.12837)\n- [表面形式竞争：为什么概率最高的答案并不总是正确的](https:\u002F\u002Farxiv.org\u002Fabs\u002F2104.08315)\n\n\u003C\u002Fdetails>\n\n### 任务分解\n\n**第一层**\n\n- [监督过程，而非结果](https:\u002F\u002Fought.org\u002Fupdates\u002F2022-04-06-process)\n- [通过放大弱专家来监督强学习者](https:\u002F\u002Farxiv.org\u002Fabs\u002F1810.08575)\n\n**第二层**\n\n- [思维之树：利用大型语言模型进行审慎的问题解决](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.10601)\n- [分解认知](https:\u002F\u002Fought.org\u002Fresearch\u002Ffactored-cognition)\n- [迭代蒸馏与放大](https:\u002F\u002Fai-alignment.com\u002Fiterated-distillation-and-amplification-157debfd1616)\n- [在人类反馈下递归总结书籍](https:\u002F\u002Farxiv.org\u002Fabs\u002F2109.10862)\n- [通过基于过程和基于结果的反馈解决数学文字题](https:\u002F\u002Farxiv.org\u002Fabs\u002F2211.14275)\n\n**第三层**\n\n- [分解验证：检测并减少学术论文摘要中的幻觉现象](https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.10627)\n- [使用大型语言模型进行忠实推理](https:\u002F\u002Farxiv.org\u002Fabs\u002F2208.14271)\n- [迭代分解：通过监督推理过程提升科学问答能力](https:\u002F\u002Farxiv.org\u002Fabs\u002F2301.01751)\n- [语言模型级联](https:\u002F\u002Farxiv.org\u002Fabs\u002F2207.10342)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>第四层及以上\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [去语境化：使句子独立存在](https:\u002F\u002Fdoi.org\u002F10.1162\u002Ftacl_a_00377)\n- [分解认知入门](https:\u002F\u002Fprimer.ought.org)\n- [思维图：利用大型语言模型解决复杂问题](http:\u002F\u002Farxiv.org\u002Fabs\u002F2308.09687)\n- [Parsel：用于算法推理的统一自然语言框架](http:\u002F\u002Farxiv.org\u002Fabs\u002F2212.10561)\n- [AI链：通过串联大型语言模型提示实现透明且可控的人机交互](https:\u002F\u002Farxiv.org\u002Fabs\u002F2110.01691)\n- [挑战BIG-Bench任务及思维链能否解决它们](https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.09261)\n- [逐项评估论证](https:\u002F\u002Fought.org\u002Fupdates\u002F2020-01-11-arguments)\n- [从最简单到最复杂提示法使大型语言模型具备复杂推理能力](https:\u002F\u002Farxiv.org\u002Fabs\u002F2205.11822)\n- [助产术式提示：借助递归解释实现逻辑一致的推理](https:\u002F\u002Farxiv.org\u002Fabs\u002F2205.11822)\n- [衡量并缩小语言模型中的组合性差距](https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.03350)\n- [PAL：程序辅助语言模型](https:\u002F\u002Farxiv.org\u002Fabs\u002F2211.10435)\n- [ReAct：在语言模型中协同推理与行动](https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.03629)\n- [选择—推理：利用大型语言模型进行可解释的逻辑推理](https:\u002F\u002Farxiv.org\u002Fabs\u002F2205.10625)\n- [展示你的工作：语言模型的中间计算草稿本](https:\u002F\u002Farxiv.org\u002Fabs\u002F2112.00114)\n- [Summ^N：针对长篇输入对话和文档的多阶段摘要框架](https:\u002F\u002Farxiv.org\u002Fabs\u002F2110.10150)\n- [Thinksum：利用大型语言模型对集合进行概率推理](https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.01293)\n\n\u003C\u002Fdetails>\n\n### 辩论\n\n**第二层**\n\n- [通过辩论实现AI安全](https:\u002F\u002Fopenai.com\u002Fblog\u002Fdebate\u002F)\n\n**第三层**\n\n- ✨ [通过证明者—估计者辩论避免混淆](https:\u002F\u002Farxiv.org\u002Fabs\u002F2506.13609)\n- ✨ [通过多智能体辩论提升语言模型的事实性和推理能力](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.14325)\n- ✨ [证明者—验证者博弈提升LLM输出的可读性](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.13692)\n- [辩论有助于监督不可靠的专家](https:\u002F\u002Ftwitter.com\u002Fjoshua_clymer\u002Fstatus\u002F1724851456967417872)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>第四层及以上\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [通过双重高效辩论实现可扩展的AI安全](http:\u002F\u002Farxiv.org\u002Fabs\u002F2311.14125)\n- [两轮辩论无法帮助人类回答困难的阅读理解问题](https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.10860)\n\n\u003C\u002Fdetails>\n\n### 工具使用与支架搭建\n\n**第二层**\n\n- [衡量训练后增强的影响](https:\u002F\u002Fmetr.github.io\u002Fautonomy-evals-guide\u002Felicitation-gap\u002F)\n- [WebGPT：在人类反馈下由浏览器辅助的问答系统](https:\u002F\u002Farxiv.org\u002Fabs\u002F2112.09332)\n\n**第三层**\n\n- ✨ [可执行代码操作能激发更好的LLM代理](https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.01030)\n- ✨ [GEPA：反思式提示进化可超越强化学习](https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.19457)\n- ✨ [TextGrad：通过文本实现自动“微分”](https:\u002F\u002Farxiv.org\u002Fabs\u002F2406.07496)\n- [无需昂贵的再训练即可显著提升AI能力](http:\u002F\u002Farxiv.org\u002Fabs\u002F2312.07413)\n- [利用语言模型自动发现统计模型](http:\u002F\u002Farxiv.org\u002Fabs\u002F2402.17879)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>第四层及以上\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [DSPy：将声明式语言模型调用编译成自我改进的流水线](http:\u002F\u002Farxiv.org\u002Fabs\u002F2310.03714)\n- [Promptbreeder：通过提示进化实现自指式的自我改进](http:\u002F\u002Farxiv.org\u002Fabs\u002F2309.16797)\n- [自学优化器（STOP）：递归式自我改进的代码生成](https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.02304v1)\n- [Voyager：一个基于大型语言模型的开放式具身智能体](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.16291)\n- [ReGAL：重构程序以发现可泛化的抽象概念](http:\u002F\u002Farxiv.org\u002Fabs\u002F2401.16467)\n\n\u003C\u002Fdetails>\n\n### 诚实、事实性和认识论\n\n**第二层**\n\n- [用于协助人类评估者的自我批判模型](https:\u002F\u002Farxiv.org\u002Fabs\u002F2206.05802v2)\n\n**第三层**\n\n- [语言模型认为哪些证据具有说服力？](http:\u002F\u002Farxiv.org\u002Fabs\u002F2402.11782)\n- [如何识破AI说谎者：通过提问无关问题在黑盒LLM中检测谎言](https:\u002F\u002Farxiv.org\u002Fabs\u002F2309.15840)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>第四层及以上\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [语言模型并不总是说出心里话：思维链提示中的不忠实解释](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.04388)\n- [大型语言模型中的长篇事实性](http:\u002F\u002Farxiv.org\u002Fabs\u002F2403.18802)\n\n\u003C\u002Fdetails>\n\n## 应用场景\n\n### 科学\n\n**第二层级**\n\n- ✨ [AlphaEvolve：用于科学和算法发现的编码智能体](https:\u002F\u002Farxiv.org\u002Fabs\u002F2506.13131)\n- ✨ [AlphaFold 3：生物分子相互作用的精确结构预测](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41586-024-07487-w)\n\n**第三层级**\n\n- ✨ [迈向人工智能联合科学家](https:\u002F\u002Fstorage.googleapis.com\u002Fdeepmind-media\u002FDeepMind.com\u002FBlog\u002Ftowards-an-ai-co-scientist\u002FTowards_an_AI_Co-Scientist.pdf)\n- [大型语言模型能否为研究论文提供有用反馈？一项大规模实证分析](http:\u002F\u002Farxiv.org\u002Fabs\u002F2310.01783)\n- [大型语言模型编码临床知识](http:\u002F\u002Farxiv.org\u002Fabs\u002F2212.13138)\n- [大型语言模型对科学发现的影响：基于GPT-4的初步研究](http:\u002F\u002Farxiv.org\u002Fabs\u002F2311.07361)\n- [以研究论文为基础的信息检索问答数据集](https:\u002F\u002Farxiv.org\u002Fabs\u002F2105.03011)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>第四层级及以上\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [通用基础模型能否超越专用微调？以医学为例的研究](http:\u002F\u002Farxiv.org\u002Fabs\u002F2311.16452)\n- [Nougat：面向学术文档的神经光学理解](http:\u002F\u002Farxiv.org\u002Fabs\u002F2308.13418)\n- [Scim：科学论文的智能略读支持](http:\u002F\u002Farxiv.org\u002Fabs\u002F2205.04561)\n- [SynerGPT：基于上下文学习的个性化药物协同效应预测与药物设计](https:\u002F\u002Fwww.biorxiv.org\u002Fcontent\u002F10.1101\u002F2023.07.06.547759v1)\n- [利用大型语言模型实现精准的鉴别诊断](http:\u002F\u002Farxiv.org\u002Fabs\u002F2312.00164)\n- [迈向人类与机器科学理解能力的基准测试](http:\u002F\u002Farxiv.org\u002Fabs\u002F2304.10327)\n- [用于发现科学挑战与研究方向的搜索引擎](https:\u002F\u002Farxiv.org\u002Fabs\u002F2108.13751)\n- [借助自动化工具，仅用两周便完成了一篇完整的系统综述：案例研究](https:\u002F\u002Fpubmed.ncbi.nlm.nih.gov\u002F32004673\u002F)\n- [事实还是虚构：科学主张的验证](https:\u002F\u002Farxiv.org\u002Fabs\u002F2004.14974)\n- [Multi-XScience：面向科学文献极端多文档摘要的大规模数据集](https:\u002F\u002Farxiv.org\u002Fabs\u002F2010.14235)\n- [PEER：协作式语言模型](https:\u002F\u002Farxiv.org\u002Fabs\u002F2208.11663)\n- [PubMedQA：生物医学研究问答数据集](https:\u002F\u002Farxiv.org\u002Fabs\u002F1909.06146)\n- [SciCo：科学概念的跨文档层次化指代消解](https:\u002F\u002Farxiv.org\u002Fabs\u002F2104.08809)\n- [SciTail：来自科学问答的文本蕴含数据集](http:\u002F\u002Fai2-website.s3.amazonaws.com\u002Fteam\u002Fashishs\u002Fscitail-aaai2018.pdf)\n\n\u003C\u002Fdetails>\n\n### 预测\n\n**第三层级**\n\n- ✨ [语言模型预测器的一致性检验](https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.18544)\n- ✨ [LLM Processes：基于自然语言条件的数值预测分布](https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.12856)\n- [AI增强预测：LLM助手提升人类预测准确性](https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.07862v1)\n- [利用语言模型逼近人类水平的预测能力](http:\u002F\u002Farxiv.org\u002Fabs\u002F2402.18563)\n- [利用神经网络预测未来世界事件](https:\u002F\u002Farxiv.org\u002Fabs\u002F2206.15474)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>第四层级及以上\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [Transformer架构在时间序列预测中是否有效？](https:\u002F\u002Farxiv.org\u002Fabs\u002F2205.13504)\n\n\u003C\u002Fdetails>\n\n### 搜索与排序\n\n**第二层级**\n\n- [大规模学习短语的稠密表示](https:\u002F\u002Farxiv.org\u002Fabs\u002F2012.12624)\n- [通过对比预训练生成文本和代码嵌入](https:\u002F\u002Farxiv.org\u002Fabs\u002F2201.10005)（OpenAI嵌入）\n\n**第三层级**\n\n- [大型语言模型通过成对排序提示可有效进行文本排序](http:\u002F\u002Farxiv.org\u002Fabs\u002F2306.17563)\n- [并非所有向量数据库都是一样的](https:\u002F\u002Fdmitry-kan.medium.com\u002Fmilvus-pinecone-vespa-weaviate-vald-gsi-what-unites-these-buzz-words-and-what-makes-each-9c65a3bd0696)\n- [REALM：检索增强型语言模型预训练](https:\u002F\u002Farxiv.org\u002Fabs\u002F2002.08909)\n- [面向知识密集型NLP任务的检索增强生成](https:\u002F\u002Farxiv.org\u002Fabs\u002F2005.11401)\n- [基于指令的任务感知检索](https:\u002F\u002Farxiv.org\u002Fabs\u002F2211.09260)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>第四层级及以上\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [RankZephyr：高效稳健的零样本列表式重排序轻而易举！](http:\u002F\u002Farxiv.org\u002Fabs\u002F2312.02724)\n- [信息检索评估中的一些常见错误及其规避方法](https:\u002F\u002Fdl.acm.org\u002Fdoi\u002F10.1145\u002F3190580.3190586)\n- [利用交互式智能体提升搜索引擎性能](https:\u002F\u002Farxiv.org\u002Fabs\u002F2109.00527)\n- [ColBERT：通过BERT上的上下文后期交互实现高效且有效的段落搜索](https:\u002F\u002Farxiv.org\u002Fabs\u002F2004.12832)\n- [超越下游任务准确率的信息检索基准测试](https:\u002F\u002Farxiv.org\u002Fabs\u002F2212.01340)\n- [UnifiedQA：以单一问答系统跨越格式边界](https:\u002F\u002Farxiv.org\u002Fabs\u002F2005.00700)\n\n\u003C\u002Fdetails>\n\n## 机器学习实践\n\n### 生产部署\n\n**第一层级**\n\n- [Python中的机器学习：数据科学、机器学习和人工智能的主要发展与技术趋势](https:\u002F\u002Farxiv.org\u002Fabs\u002F2002.04803v2)\n- [机器学习：技术债务中的高息信用卡](https:\u002F\u002Fpapers.nips.cc\u002Fpaper\u002F2015\u002Ffile\u002F86df7dcfd896fcaf2674f757a2463eba-Paper.pdf)\n\n**第二层级**\n\n- [设计数据密集型应用](https:\u002F\u002Fdataintensive.net\u002F)\n- [训练神经网络的配方](http:\u002F\u002Fkarpathy.github.io\u002F2019\u002F04\u002F25\u002Frecipe\u002F)（Karpathy）\n\n### 基准测试\n\n**第二层级**\n\n- ✨ [GAIA：通用人工智能助手的基准测试](http:\u002F\u002Farxiv.org\u002Fabs\u002F2311.12983)\n- [GPQA：一项研究生级别的防谷歌问答基准测试](http:\u002F\u002Farxiv.org\u002Fabs\u002F2311.12022)\n- [SWE-bench：语言模型能否解决现实世界的 GitHub 问题？](https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.06770v1)\n- [TruthfulQA：衡量模型如何模仿人类的谬误](https:\u002F\u002Farxiv.org\u002Fabs\u002F2109.07958)\n\n**第三层级**\n\n- ✨ [RE-Bench：评估前沿人工智能研发能力](https:\u002F\u002Farxiv.org\u002Fabs\u002F2411.15114)\n- ✨ [SimpleQA：衡量简短形式的事实准确性](https:\u002F\u002Fopenai.com\u002Findex\u002Fintroducing-simpleqa\u002F)\n- ✨ [ARC Prize 2024：技术报告](https:\u002F\u002Farcprize.org\u002Fblog\u002Foai-o3-pub-breakthrough)\n- ✨ [FrontierMath：评估人工智能高级数学推理能力的基准测试](https:\u002F\u002Farxiv.org\u002Fabs\u002F2411.04872)\n- [衡量大规模多任务语言理解能力](https:\u002F\u002Farxiv.org\u002Fabs\u002F2009.03300)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>第四层级及以上\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [FLEX：统一的小样本 NLP 评估框架](https:\u002F\u002Farxiv.org\u002Fabs\u002F2107.07170)\n- [语言模型的整体评估](https:\u002F\u002Farxiv.org\u002Fabs\u002F2107.07170)（HELM）\n- [使用语言模型实现真正的少样本学习](https:\u002F\u002Farxiv.org\u002Fabs\u002F2105.11447)\n- [ConditionalQA：一个具有条件答案的复杂阅读理解数据集](https:\u002F\u002Farxiv.org\u002Fabs\u002F2110.06884)\n- [使用 MATH 数据集衡量数学问题解决能力](https:\u002F\u002Farxiv.org\u002Fabs\u002F2103.03874)\n- [QuALITY：长篇输入文本的问答，当然可以！](https:\u002F\u002Farxiv.org\u002Fabs\u002F2112.08608)\n- [SCROLLS：针对长序列语言的标准比较框架](https:\u002F\u002Farxiv.org\u002Fabs\u002F2201.03533)\n- [自然语言理解领域的基准测试需要怎样的改进？](https:\u002F\u002Farxiv.org\u002Fabs\u002F2104.02145)\n\n\u003C\u002Fdetails>\n\n### 数据集\n\n**第二层级**\n\n- [Common Crawl](https:\u002F\u002Farxiv.org\u002Fabs\u002F2105.02732)\n- [The Pile：用于语言建模的 800GB 多样化文本数据集](https:\u002F\u002Farxiv.org\u002Fabs\u002F2101.00027)\n\n**第三层级**\n\n- ✨ [FineWeb：大规模精选优质网络文本数据](https:\u002F\u002Fhuggingface.co\u002Fspaces\u002FHuggingFaceFW\u002Fblogpost-fineweb-v1)\n- [对话修复：将文档转化为对话](https:\u002F\u002Farxiv.org\u002Fabs\u002F2205.09073)\n- [MS MARCO：一个人工生成的机器阅读理解数据集](https:\u002F\u002Farxiv.org\u002Fabs\u002F1611.09268)\n- [微软学术图谱](https:\u002F\u002Finternal-journal.frontiersin.org\u002Farticles\u002F10.3389\u002Ffdata.2019.00045\u002Ffull)\n\n## 高级主题\n\n### 世界模型与因果关系\n\n**第三层级**\n\n- [涌现的世界表征：探索在合成任务上训练的序列模型](http:\u002F\u002Farxiv.org\u002Fabs\u002F2210.13382)\n- [从词模型到世界模型：从自然语言到概率性思维语言的转换](http:\u002F\u002Farxiv.org\u002Fabs\u002F2306.12672)\n- [语言模型对空间和时间的表征](http:\u002F\u002Farxiv.org\u002Fabs\u002F2310.02207)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>第四层级及以上\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [大型语言模型中难以处理的推理的摊销](http:\u002F\u002Farxiv.org\u002Fabs\u002F2310.04363)\n- [CLADDER：评估语言模型中的因果推理能力](http:\u002F\u002Fzhijing-jin.com\u002Ffiles\u002Fpapers\u002FCLadder_2023.pdf)\n- [因果贝叶斯优化](https:\u002F\u002Fproceedings.mlr.press\u002Fv108\u002Faglietti20a.html)\n- [因果推理与大型语言模型：开启因果关系的新前沿](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.00050)\n- [生成式智能体：人类行为的交互式模拟](http:\u002F\u002Farxiv.org\u002Fabs\u002F2304.03442)\n- [智能体和语言模型中主动因果策略的被动学习](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.16183)\n\n\u003C\u002Fdetails>\n\n### 规划\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>第四层级及以上\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [超越 A\\*：通过搜索动态自举实现更好的规划](http:\u002F\u002Farxiv.org\u002Fabs\u002F2402.14083)\n- [面向语言智能体的认知架构](http:\u002F\u002Farxiv.org\u002Fabs\u002F2309.02427)\n\n\u003C\u002Fdetails>\n\n### 不确定性、校准与主动学习\n\n**第二层级**\n\n- [专家不会作弊：通过预测配对来学习未知知识](http:\u002F\u002Farxiv.org\u002Fabs\u002F2402.08733)\n- [深度学习中贝叶斯不确定性的简单基线](https:\u002F\u002Farxiv.org\u002Fabs\u002F1902.02476)\n- [Plex：利用预训练的大模型扩展提升可靠性](https:\u002F\u002Farxiv.org\u002Fabs\u002F2207.07411)\n\n**第三层级**\n\n- ✨ [文本贝叶斯：量化基于 LLM 的系统的不确定性](https:\u002F\u002Farxiv.org\u002Fabs\u002F2506.10060)\n- [利用语言模型和概率推理进行主动偏好推断](http:\u002F\u002Farxiv.org\u002Fabs\u002F2312.12009)\n- [用语言模型引出人类偏好](http:\u002F\u002Farxiv.org\u002Fabs\u002F2310.11589)\n- [用自然语言描述文本分布之间的差异](https:\u002F\u002Farxiv.org\u002Fabs\u002F2201.12323)\n- [教会模型用语言表达其不确定性](https:\u002F\u002Farxiv.org\u002Fabs\u002F2205.14334)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>第四层级及以上\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [通过获取对比示例进行主动学习](https:\u002F\u002Farxiv.org\u002Fabs\u002F2109.03764)\n- [使用自然语言和概率推理进行实验并修订规则](http:\u002F\u002Farxiv.org\u002Fabs\u002F2402.06025)\n- [STaR-GATE：教会语言模型提出澄清问题](http:\u002F\u002Farxiv.org\u002Fabs\u002F2403.19154)\n- [主动测试：高效的模型评估](https:\u002F\u002Farxiv.org\u002Fabs\u002F2103.05331)\n- [语言奖励模型的不确定性估计](https:\u002F\u002Farxiv.org\u002Fabs\u002F2203.07472)\n\n\u003C\u002Fdetails>\n\n### 可解释性与模型编辑\n\n**2级**\n\n- ✨ [单义性扩展：从 Claude 3 Sonnet 中提取可解释特征](https:\u002F\u002Ftransformer-circuits.pub\u002F2024\u002Fscaling-monosemanticity\u002F)\n- ✨ [规模化可解释性：识别 Alpaca 中的因果机制](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.08809)\n- [在无监督条件下发现语言模型中的潜在知识](https:\u002F\u002Farxiv.org\u002Fabs\u002F2212.03827v1)\n\n**3级**\n\n- ✨ [稀疏自编码器的扩展与评估](https:\u002F\u002Fcdn.openai.com\u002Fpapers\u002Fsparse-autoencoders.pdf)\n- ✨ [打开 AI 黑箱：通过机制性可解释性进行程序合成](https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.05110v1)\n- [从机制角度分析微调对过程化定义任务的影响](http:\u002F\u002Farxiv.org\u002Fabs\u002F2311.12786)\n- [表征工程：一种自上而下的 AI 透明度方法](http:\u002F\u002Farxiv.org\u002Fabs\u002F2310.01405)\n- [利用影响函数研究大型语言模型的泛化能力](http:\u002F\u002Farxiv.org\u002Fabs\u002F2308.03296)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>4级及以上\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [码本特征：神经网络的稀疏离散可解释性](http:\u002F\u002Farxiv.org\u002Fabs\u002F2310.17230)\n- [借助调谐透镜从 Transformer 中提取潜在预测](http:\u002F\u002Farxiv.org\u002Fabs\u002F2303.08112)\n- [语言模型如何在上下文中绑定实体？](http:\u002F\u002Farxiv.org\u002Fabs\u002F2310.17191)\n- [野外的可解释性：GPT-2 small 中用于间接宾语识别的电路](https:\u002F\u002Farxiv.org\u002Fabs\u002F2211.00593)\n- [稀疏特征电路：在语言模型中发现并编辑可解释的因果图](http:\u002F\u002Farxiv.org\u002Fabs\u002F2403.19647)\n- [揭示 Transformer 中的内生优化算法](http:\u002F\u002Farxiv.org\u002Fabs\u002F2309.05858)\n- [大规模快速模型编辑](https:\u002F\u002Farxiv.org\u002Fabs\u002F2110.11309)\n- [Git Re-Basin：在置换对称性下合并模型](https:\u002F\u002Farxiv.org\u002Fabs\u002F2209.04836)\n- [定位并编辑 GPT 中的事实关联](https:\u002F\u002Farxiv.org\u002Fabs\u002F2202.05262)\n- [在 Transformer 中批量编辑记忆](https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.07229)\n\n\u003C\u002Fdetails>\n\n### 强化学习\n\n**2级**\n\n- ✨ [DeepSeekMath：突破开放语言模型的数学推理极限](https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.03300) (GRPO)\n- [直接偏好优化：你的语言模型其实是个奖励模型](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.18290)\n- [Reflexion：具备口头强化学习能力的语言智能体](http:\u002F\u002Farxiv.org\u002Fabs\u002F2303.11366)\n- [使用通用强化学习算法通过自我博弈掌握国际象棋和将棋](https:\u002F\u002Farxiv.org\u002Fabs\u002F1712.01815) (AlphaZero)\n- [MuZero：通过基于学习模型的规划掌握雅达利、围棋、国际象棋和将棋](https:\u002F\u002Farxiv.org\u002Fabs\u002F1911.08265)\n\n**3级**\n\n- [人类反馈强化学习的开放问题与根本局限](http:\u002F\u002Farxiv.org\u002Fabs\u002F2307.15217)\n- [AlphaStar：掌握即时战略游戏《星际争霸 II》](https:\u002F\u002Fdeepmind.com\u002Fblog\u002Farticle\u002Falphastar-mastering-real-time-strategy-game-starcraft-ii)\n- [决策 Transformer](https:\u002F\u002Farxiv.org\u002Fabs\u002F2106.01345)\n- [仅用少量数据掌握雅达利游戏](https:\u002F\u002Farxiv.org\u002Fabs\u002F2111.00210) (EfficientZero)\n- [掌握经典不完全信息博弈 Stratego](https:\u002F\u002Fwww.science.org\u002Fdoi\u002F10.1126\u002Fscience.add4679) (DeepNash)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>4级及以上\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [AlphaStar Unplugged：大规模离线强化学习](http:\u002F\u002Farxiv.org\u002Fabs\u002F2308.03526)\n- [有限认知负荷下的贝叶斯强化学习](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.03263)\n- [对比偏好学习：无需强化学习即可从人类反馈中学习](http:\u002F\u002Farxiv.org\u002Fabs\u002F2310.13639)\n- [无需搜索即可达到特级大师级别的国际象棋水平](http:\u002F\u002Farxiv.org\u002Fabs\u002F2402.04494)\n- [一种数据驱动的学习控制计算机的方法](https:\u002F\u002Farxiv.org\u002Fabs\u002F2202.08137)\n- [AlphaZero 中国际象棋知识的习得](https:\u002F\u002Farxiv.org\u002Fabs\u002F2111.09259)\n- [游戏玩家](https:\u002F\u002Farxiv.org\u002Fabs\u002F2112.03178)\n- [检索增强型强化学习](https:\u002F\u002Farxiv.org\u002Fabs\u002F2202.08417)\n\n\u003C\u002Fdetails>\n\n## 全局视角\n\n### AI 扩展\n\n**1级**\n\n- [神经语言模型的规模法则](https:\u002F\u002Farxiv.org\u002Fabs\u002F2001.08361)\n- [起飞速度](https:\u002F\u002Fsideways-view.com\u002F2018\u002F02\u002F24\u002Ftakeoff-speeds\u002F)\n- [苦涩的教训](http:\u002F\u002Fwww.incompleteideas.net\u002FIncIdeas\u002FBitterLesson.html)\n\n**2级**\n\n- [AI 与计算资源](https:\u002F\u002Fopenai.com\u002Fblog\u002Fai-and-compute\u002F)\n- [迁移学习的规模法则](https:\u002F\u002Farxiv.org\u002Fabs\u002F2102.01293)\n- [训练计算最优的大语言模型](https:\u002F\u002Farxiv.org\u002Fabs\u002F2203.15556) (Chinchilla)\n\n**3级**\n\n- ✨ [无限计算条件下的预训练](https:\u002F\u002Farxiv.org\u002Fabs\u002F2509.14786)\n- [大型语言模型的涌现能力](https:\u002F\u002Farxiv.org\u002Fabs\u002F2206.07682)\n- [以 0.1% 的额外计算超越规模法则](https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.11399) (U-PaLM)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>4级及以上\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [语言模型物理学：第 3.3 部分，知识容量规模法则](http:\u002F\u002Farxiv.org\u002Fabs\u002F2404.05405)\n- [速通与机器学习中的幂律趋势](http:\u002F\u002Farxiv.org\u002Fabs\u002F2304.10004)\n- [单智能体强化学习的规模法则](http:\u002F\u002Farxiv.org\u002Fabs\u002F2301.13442)\n- [超越神经网络规模法则：通过数据剪枝打破幂律缩放](https:\u002F\u002Farxiv.org\u002Fabs\u002F2206.14486)\n- [大型语言模型的涌现能力](https:\u002F\u002Farxiv.org\u002Fabs\u002F2206.07682)\n- [用棋类游戏扩展规模法则](https:\u002F\u002Farxiv.org\u002Fabs\u002F2104.03113)\n\n\u003C\u002Fdetails>\n\n### 人工智能安全\n\n**一级**\n\n- [机器智能的三大影响](https:\u002F\u002Fwww.effectivealtruism.org\u002Farticles\u002Fthree-impacts-of-machine-intelligence-paul-christiano\u002F)\n- [失败的样子](https:\u002F\u002Fwww.alignmentforum.org\u002Fposts\u002FHBxe6wdjxK239zajf\u002Fwhat-failure-looks-like)\n- [如果没有具体的应对措施，通往变革性人工智能的最简单路径很可能导致AI接管](https:\u002F\u002Fwww.lesswrong.com\u002Fposts\u002FpRkFkzwKZ2zfa3R6H\u002Fwithout-specific-countermeasures-the-easiest-path-to)\n\n**二级**\n\n- [灾难性AI风险概述](http:\u002F\u002Farxiv.org\u002Fabs\u002F2306.12001)\n- [澄清“失败的样子”（第1部分）](https:\u002F\u002Fwww.lesswrong.com\u002Fposts\u002Fv6Q7T335KCMxujhZu\u002Fclarifying-what-failure-looks-like-part-1)\n- [基于人类偏好进行深度强化学习](https:\u002F\u002Fopenai.com\u002Fblog\u002Fdeep-reinforcement-learning-from-human-preferences\u002F)\n- [从深度学习角度看对齐问题](https:\u002F\u002Farxiv.org\u002Fabs\u002F2209.00626)\n\n**三级**\n\n- ✨ [大型语言模型中的对齐伪装](https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.14093)\n- ✨ [宪法分类器：防御通用越狱攻击](https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.18837)\n- ✨ [涌现式不对齐：窄范围微调可能导致广泛不对齐的语言模型](https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.17424)\n- ✨ [逐步去权力化：渐进式AI发展带来的系统性生存风险](https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.16946)\n- [阴谋型AI：AI会在训练过程中假装对齐以获取权力吗？](http:\u002F\u002Farxiv.org\u002Fabs\u002F2311.08379)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>四级及以上\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [面向启发式估计器的迭代期望法则探索](https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.01290)\n- [衡量大型语言模型可扩展监督的进展](https:\u002F\u002Farxiv.org\u002Fabs\u002F2211.03540)\n- [通过奖励建模实现可扩展的智能体对齐](https:\u002F\u002Farxiv.org\u002Fabs\u002F1811.07871)\n- [AI欺骗：案例、风险及潜在解决方案综述](http:\u002F\u002Farxiv.org\u002Fabs\u002F2308.14752)\n- [检测测量篡改的基准测试](http:\u002F\u002Farxiv.org\u002Fabs\u002F2308.15605)\n- [国际象棋作为AI安全“预言家”方法的试验场](http:\u002F\u002Farxiv.org\u002Fabs\u002F2010.02911)\n- [关闭通向非人未来的大门：我们为何以及如何选择不开发超人类通用人工智能](https:\u002F\u002Fpapers.ssrn.com\u002Fabstract=4608505)\n- [极端风险下的模型评估](http:\u002F\u002Farxiv.org\u002Fabs\u002F2305.15324)\n- [前沿AI发展的负责任报道](http:\u002F\u002Farxiv.org\u002Fabs\u002F2404.02675)\n- [安全论证：如何证明先进AI系统的安全性](http:\u002F\u002Farxiv.org\u002Fabs\u002F2403.10462)\n- [休眠代理：训练能在安全训练中持续存在的欺骗性语言模型](http:\u002F\u002Farxiv.org\u002Fabs\u002F2401.05566)\n- [技术报告：大型语言模型在压力下会策略性地欺骗用户](http:\u002F\u002Farxiv.org\u002Fabs\u002F2311.07590)\n- [张量信任：来自在线游戏的可解释提示注入攻击](http:\u002F\u002Farxiv.org\u002Fabs\u002F2311.01011)\n- [验证神经网络训练数据的工具](http:\u002F\u002Farxiv.org\u002Fabs\u002F2307.00682)\n- [迈向具有收敛安全边界谨慎科学家AI的探索](https:\u002F\u002Fyoshuabengio.org\u002F2024\u002F02\u002F26\u002Ftowards-a-cautious-scientist-ai-with-convergent-safety-bounds\u002F)\n- [语言智能体的对齐](https:\u002F\u002Farxiv.org\u002Fabs\u002F2103.14659)\n- [激发潜在知识](https:\u002F\u002Fdocs.google.com\u002Fdocument\u002Fd\u002F1WwsnJQstPq91_Yh-Ch2XRL8H_EpsnjrC1dwZXR37PC8\u002Fedit?usp=sharing)\n- [使用语言模型对抗语言模型以减少危害：方法、规模化行为与经验教训](https:\u002F\u002Farxiv.org\u002Fabs\u002F2209.07858)\n- [用语言模型对抗语言模型](https:\u002F\u002Fstorage.googleapis.com\u002Fdeepmind-media\u002FRed%20Teaming\u002FRed%20Teaming.pdf)\n- [高级机器学习系统中学习优化带来的风险](https:\u002F\u002Farxiv.org\u002Fabs\u002F1906.01820)\n- [ML安全领域的未解决问题](https:\u002F\u002Farxiv.org\u002Fabs\u002F2109.13916)\n\n\u003C\u002Fdetails>\n\n### 经济与社会影响\n\n**二级**\n\n- ✨ [AI 2027](https:\u002F\u002Fai-2027.com\u002F)\n- ✨ [情境感知](https:\u002F\u002Fsituational-awareness.ai\u002F) (Aschenbrenner)\n\n**三级**\n\n- [AI自动化带来的爆炸性增长：论点回顾](http:\u002F\u002Farxiv.org\u002Fabs\u002F2309.11690)\n- [语言模型可以减少信息市场的不对称性](http:\u002F\u002Farxiv.org\u002Fabs\u002F2403.14443)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>四级及以上\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [弥合人机知识鸿沟：AlphaZero中的概念发现与迁移](http:\u002F\u002Farxiv.org\u002Fabs\u002F2310.16410)\n- [基础模型与合理使用](https:\u002F\u002Farxiv.org\u002Fabs\u002F2303.15715v1)\n- [GPT就是GPT：大型语言模型对劳动力市场影响潜力的早期观察](http:\u002F\u002Farxiv.org\u002Fabs\u002F2303.10130)\n- [AGI的层次：将通往AGI的进展具体化](http:\u002F\u002Farxiv.org\u002Fabs\u002F2311.02462)\n- [LLM用于与Polis进行可扩展审议的机会与风险](http:\u002F\u002Farxiv.org\u002Fabs\u002F2306.11932)\n- [关于基础模型的机会与风险](https:\u002F\u002Farxiv.org\u002Fabs\u002F2108.07258)\n\n\u003C\u002Fdetails>\n\n### 哲学\n\n**二级**\n\n- [大型语言模型中的无指称意义](https:\u002F\u002Farxiv.org\u002Fabs\u002F2208.02957)\n\n\u003Cdetails>\u003Csummary>\u003Cstrong>四级及以上\u003C\u002Fstrong>\u003C\u002Fsummary>\n\n- [人工智能中的意识：来自意识科学的洞见](http:\u002F\u002Farxiv.org\u002Fabs\u002F2308.08708)\n- [哲学家应当开发、理论化并使用具有哲学相关性的AI](https:\u002F\u002Fphilarchive.org\u002Farchive\u002FCLAPOT-16)\n- [迈向利用自我报告评估AI系统道德地位的方法](http:\u002F\u002Farxiv.org\u002Fabs\u002F2311.08576)\n\n\u003C\u002Fdetails>\n\n## 维护者\n\n[andreas@elicit.com](mailto:andreas@elicit.com)","# machine-learning-list 快速上手指南\n\n## 简介\n`machine-learning-list` 并非一个可安装的软件库或框架，而是一份由 Elicit 团队维护的**机器学习与大语言模型（LLM）精选阅读清单**。它旨在帮助开发者系统性地学习从基础理论到生产部署的核心知识。本指南将指导你如何获取并利用这份资源进行学习。\n\n## 环境准备\n\n由于本项目本质上是文档列表，无需特定的操作系统或复杂的依赖环境。你只需要：\n\n*   **操作系统**：任意支持现代浏览器的系统（Windows, macOS, Linux）。\n*   **前置依赖**：\n    *   稳定的网络连接（部分论文链接可能需要学术网络环境）。\n    *   Git（可选，用于克隆仓库本地阅读）。\n    *   Markdown 阅读器或直接使用 GitHub 网页版。\n\n## 安装步骤\n\n你可以通过以下两种方式获取该阅读清单：\n\n### 方式一：在线直接阅读（推荐）\n直接访问 GitHub 仓库页面，无需安装任何工具。\n*   地址：[https:\u002F\u002Fgithub.com\u002Felicit\u002Fmachine-learning-list](https:\u002F\u002Fgithub.com\u002Felicit\u002Fmachine-learning-list)\n\n### 方式二：克隆到本地\n如果你希望离线阅读或通过代码编辑器管理学习进度，可以使用 Git 克隆：\n\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002Felicit\u002Fmachine-learning-list.git\ncd machine-learning-list\n```\n\n若在国内访问 GitHub 较慢，可使用国内镜像源加速克隆：\n\n```bash\ngit clone https:\u002F\u002Fgitee.com\u002Fmirrors\u002Fmachine-learning-list.git\n# 注意：如果 Gitee 镜像未同步最新内容，请优先使用官方源或配置 Git 代理\n```\n\n## 基本使用\n\n本项目的核心使用方法是按照推荐的**分级阅读顺序**进行学习。\n\n### 1. 遵循阅读层级\n清单将每个主题下的资料分为不同等级（Tier），建议严格按照以下顺序阅读：\n*   **Tier 1**：核心基础，必读内容（涵盖直观介绍和关键概念）。\n*   **Tier 2**：进阶深入，包含关键技术细节和经典论文。\n*   **Tier 3+**：高级专题，适合特定领域的深入研究。\n\n### 2. 学习路径示例\n假设你想学习 **Transformers** 架构，请在 `README.md` 中找到对应章节：\n\n1.  **第一步（Tier 1）**：观看视频《But what is a GPT?》和阅读《The Illustrated Transformer》，建立直观认知。\n2.  **第二步（Tier 2）**：阅读原始论文《Attention Is All You Need》及 Karpathy 的深度解析视频，理解技术细节。\n3.  **第三步（Tier 3）**：研读《The Annotated Transformer》代码实现或最新的研究论文（如带有 ✨ 标记的 2025 年后新文）。\n\n### 3. 利用目录导航\n项目涵盖了从 [基础 (Fundamentals)](#fundamentals) 到 [宏观视野 (The big picture)](#the-big-picture) 的七大板块。你可以直接点击目录跳转到感兴趣的主题，例如：\n*   **推理策略**：学习 In-context reasoning 和 Task decomposition。\n*   **实战应用**：查看 Production deployment 和 Benchmarks 相关文献。\n*   **训练微调**：参考 LoRA、QLoRA 等高效微调技术的论文。\n\n> **提示**：列表中标记为 `✨` 的条目表示为 2025 年 11 月 26 日之后新增的前沿内容，建议重点关注以获取最新技术动态。","某科技公司的新晋 AI 工程师团队正急需构建大语言模型应用，但成员背景各异，对从基础理论到前沿部署的知识体系缺乏统一认知。\n\n### 没有 machine-learning-list 时\n- **学习路径混乱**：团队成员在海量论文和教程中盲目摸索，有人沉迷过时的反向推导数学细节，有人直接跳跃阅读高深架构，导致知识断层严重。\n- **理论与实战脱节**：大家花费大量时间研读纯学术理论，却忽略了“生产环境部署”和“基准测试”等关键工程环节，模型无法落地。\n- **前沿视野缺失**：由于缺乏系统指引，团队对“上下文推理”、“工具使用”及\"AI 安全”等决定产品竞争力的前沿策略知之甚少。\n- **沟通成本高昂**：因缺乏共同的知识基准，技术评审时大家对基础概念理解不一，反复解释基础术语浪费了宝贵的开发时间。\n\n### 使用 machine-learning-list 后\n- **路径清晰高效**：团队严格遵循 Tier 1 至 Tier 3 的分级阅读顺序，先通过 Karpathy 的视频直观掌握 Transformer 核心，再深入微积分细节，全员快速对齐基础。\n- **工程导向明确**：课程专门涵盖“生产部署”与“数据集”章节，引导成员在学习初期就关注模型在实际业务中的可扩展性与稳定性。\n- **掌握前沿策略**：通过“推理与运行时策略”模块，团队迅速掌握了任务分解、辩论机制及工具调用等高级技巧，显著提升了模型解决复杂问题的能力。\n- **协作无缝顺畅**：所有人基于同一份权威大纲建立知识体系，技术讨论时术语统一、逻辑同频，大幅缩短了从学习到编码的转化周期。\n\nmachine-learning-list 通过将碎片化的机器学习知识重构为从入门到前沿的系统化课程，帮助团队以最低成本建立了兼具理论深度与工程广度的核心竞争力。","https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Felicit_machine-learning-list_6dad6f1f.png","elicit","Elicit","https:\u002F\u002Foss.gittoolsai.com\u002Favatars\u002Felicit_43406f6e.png","AI for Scientific Research",null,"elicitorg","https:\u002F\u002Felicit.com\u002F","https:\u002F\u002Fgithub.com\u002Felicit",1448,125,"2026-04-04T21:35:34",1,"","未说明",{"notes":88,"python":86,"dependencies":89},"该工具并非可执行的软件代码库，而是一份机器学习阅读清单（Curriculum\u002FReading List）。它由指向论文、博客文章和视频的链接组成，旨在帮助员工学习机器学习背景知识。因此，该工具本身没有操作系统、GPU、内存、Python 版本或依赖库的安装需求。用户只需使用网页浏览器访问列出的链接即可。",[],[14,35],[92,93,94,95],"artificial-intelligence","language-model","machine-learning","transformers","2026-03-27T02:49:30.150509","2026-04-07T22:49:55.203448",[],[]]