[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"similar-sierra-research--tau2-bench":3,"tool-sierra-research--tau2-bench":64},[4,17,27,35,43,56],{"id":5,"name":6,"github_repo":7,"description_zh":8,"stars":9,"difficulty_score":10,"last_commit_at":11,"category_tags":12,"status":16},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。",162132,3,"2026-04-05T11:01:52",[13,14,15],"开发框架","图像","Agent","ready",{"id":18,"name":19,"github_repo":20,"description_zh":21,"stars":22,"difficulty_score":23,"last_commit_at":24,"category_tags":25,"status":16},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",138956,2,"2026-04-05T11:33:21",[13,15,26],"语言模型",{"id":28,"name":29,"github_repo":30,"description_zh":31,"stars":32,"difficulty_score":23,"last_commit_at":33,"category_tags":34,"status":16},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。",107662,"2026-04-03T11:11:01",[13,14,15],{"id":36,"name":37,"github_repo":38,"description_zh":39,"stars":40,"difficulty_score":23,"last_commit_at":41,"category_tags":42,"status":16},3704,"NextChat","ChatGPTNextWeb\u002FNextChat","NextChat 是一款轻量且极速的 AI 助手，旨在为用户提供流畅、跨平台的大模型交互体验。它完美解决了用户在多设备间切换时难以保持对话连续性，以及面对众多 AI 模型不知如何统一管理的痛点。无论是日常办公、学习辅助还是创意激发，NextChat 都能让用户随时随地通过网页、iOS、Android、Windows、MacOS 或 Linux 端无缝接入智能服务。\n\n这款工具非常适合普通用户、学生、职场人士以及需要私有化部署的企业团队使用。对于开发者而言，它也提供了便捷的自托管方案，支持一键部署到 Vercel 或 Zeabur 等平台。\n\nNextChat 的核心亮点在于其广泛的模型兼容性，原生支持 Claude、DeepSeek、GPT-4 及 Gemini Pro 等主流大模型，让用户在一个界面即可自由切换不同 AI 能力。此外，它还率先支持 MCP（Model Context Protocol）协议，增强了上下文处理能力。针对企业用户，NextChat 提供专业版解决方案，具备品牌定制、细粒度权限控制、内部知识库整合及安全审计等功能，满足公司对数据隐私和个性化管理的高标准要求。",87618,"2026-04-05T07:20:52",[13,26],{"id":44,"name":45,"github_repo":46,"description_zh":47,"stars":48,"difficulty_score":23,"last_commit_at":49,"category_tags":50,"status":16},2268,"ML-For-Beginners","microsoft\u002FML-For-Beginners","ML-For-Beginners 是由微软推出的一套系统化机器学习入门课程，旨在帮助零基础用户轻松掌握经典机器学习知识。这套课程将学习路径规划为 12 周，包含 26 节精炼课程和 52 道配套测验，内容涵盖从基础概念到实际应用的完整流程，有效解决了初学者面对庞大知识体系时无从下手、缺乏结构化指导的痛点。\n\n无论是希望转型的开发者、需要补充算法背景的研究人员，还是对人工智能充满好奇的普通爱好者，都能从中受益。课程不仅提供了清晰的理论讲解，还强调动手实践，让用户在循序渐进中建立扎实的技能基础。其独特的亮点在于强大的多语言支持，通过自动化机制提供了包括简体中文在内的 50 多种语言版本，极大地降低了全球不同背景用户的学习门槛。此外，项目采用开源协作模式，社区活跃且内容持续更新，确保学习者能获取前沿且准确的技术资讯。如果你正寻找一条清晰、友好且专业的机器学习入门之路，ML-For-Beginners 将是理想的起点。",84991,"2026-04-05T10:45:23",[14,51,52,53,15,54,26,13,55],"数据工具","视频","插件","其他","音频",{"id":57,"name":58,"github_repo":59,"description_zh":60,"stars":61,"difficulty_score":10,"last_commit_at":62,"category_tags":63,"status":16},3128,"ragflow","infiniflow\u002Fragflow","RAGFlow 是一款领先的开源检索增强生成（RAG）引擎，旨在为大语言模型构建更精准、可靠的上下文层。它巧妙地将前沿的 RAG 技术与智能体（Agent）能力相结合，不仅支持从各类文档中高效提取知识，还能让模型基于这些知识进行逻辑推理和任务执行。\n\n在大模型应用中，幻觉问题和知识滞后是常见痛点。RAGFlow 通过深度解析复杂文档结构（如表格、图表及混合排版），显著提升了信息检索的准确度，从而有效减少模型“胡编乱造”的现象，确保回答既有据可依又具备时效性。其内置的智能体机制更进一步，使系统不仅能回答问题，还能自主规划步骤解决复杂问题。\n\n这款工具特别适合开发者、企业技术团队以及 AI 研究人员使用。无论是希望快速搭建私有知识库问答系统，还是致力于探索大模型在垂直领域落地的创新者，都能从中受益。RAGFlow 提供了可视化的工作流编排界面和灵活的 API 接口，既降低了非算法背景用户的上手门槛，也满足了专业开发者对系统深度定制的需求。作为基于 Apache 2.0 协议开源的项目，它正成为连接通用大模型与行业专有知识之间的重要桥梁。",77062,"2026-04-04T04:44:48",[15,14,13,26,54],{"id":65,"github_repo":66,"name":67,"description_en":68,"description_zh":69,"ai_summary_zh":69,"readme_en":70,"readme_zh":71,"quickstart_zh":72,"use_case_zh":73,"hero_image_url":74,"owner_login":75,"owner_name":76,"owner_avatar_url":77,"owner_bio":78,"owner_company":79,"owner_location":79,"owner_email":79,"owner_twitter":80,"owner_website":81,"owner_url":82,"languages":83,"stars":111,"forks":112,"last_commit_at":113,"license":114,"difficulty_score":23,"env_os":115,"env_gpu":116,"env_ram":116,"env_deps":117,"category_tags":129,"github_topics":130,"view_count":23,"oss_zip_url":79,"oss_zip_packed_at":79,"status":16,"created_at":136,"updated_at":137,"faqs":138,"releases":168},2324,"sierra-research\u002Ftau2-bench","tau2-bench","τ-Bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains","tau2-bench 是一个专为评估现实世界中“工具 - 智能体 - 用户”交互性能而设计的仿真基准框架。它主要解决了当前 AI 客服智能体在复杂业务场景下缺乏标准化、高质量评估手段的难题，帮助开发者量化智能体在遵循业务政策、调用外部工具以及处理多轮对话时的真实表现。\n\n该工具非常适合 AI 研究人员、大模型应用开发者以及企业级智能体构建者使用。通过模拟航空、零售、电信及银行业务等多种真实领域，tau2-bench 能够测试智能体是否能在严格约束下准确完成任务。其独特的技术亮点在于支持从传统的文本半双工（轮流对话）到先进的语音全双工（实时双向通话）评估模式，并集成了基于检索增强生成（RAG）的知识库查询能力。此外，最新版本还修复了大量任务逻辑缺陷，提供了包含实时语音交互和知识检索的综合排行榜，为优化智能体的协作能力提供了可靠的数据支撑。","# $\\tau$-Bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains\n\n[![python](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FPython-3.12%2B-blue.svg?style=flat&logo=python&logoColor=white)](https:\u002F\u002Fwww.python.org)\n[![Ruff](https:\u002F\u002Fimg.shields.io\u002Fendpoint?url=https:\u002F\u002Fraw.githubusercontent.com\u002Fastral-sh\u002Fruff\u002Fmain\u002Fassets\u002Fbadge\u002Fv2.json)](https:\u002F\u002Fgithub.com\u002Fastral-sh\u002Fruff)\n[![arXiv](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fcs.AI-arXiv%3A2506.07982-B31B1B.svg?logo=arxiv&logoColor=red)](https:\u002F\u002Farxiv.org\u002Fabs\u002F2506.07982)\n[![blog](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fblog-tau--bench-green)](https:\u002F\u002Fsierra.ai\u002Fblog\u002Fbenchmarking-agents-in-collaborative-real-world-scenarios)\n[![Twitter](https:\u002F\u002Fimg.shields.io\u002Ftwitter\u002Furl\u002Fhttps\u002Ftwitter.com\u002Fsierra.svg?style=social&label=Follow%20%40SierraPlatform)](https:\u002F\u002Fx.com\u002FSierraPlatform\u002Fstatus\u002F1932464265207889974)\n[![LinkedIn](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLinkedIn-0077B5?logo=linkedin&logoColor=white)](https:\u002F\u002Fwww.linkedin.com\u002Fposts\u002Fsierra_last-year-we-introduced-%F0%9D%9C%8F-bench-a-benchmark-activity-7338229693898231809-F8L4?utm_source=share&utm_medium=member_desktop&rcm=ACoAAAdc8goBmhEsiEo1_t_XSJbAnY4_zMfAWcE)\n[![Leaderboard](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F🏆_Live_Leaderboard-taubench.com-brightgreen?style=flat)](https:\u002F\u002Ftaubench.com)\n\n\u003Cdiv align=\"center\">\n\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fsierra-research_tau2-bench_readme_018677adf914.png\" width=\"95%\" alt=\"Trajectory\">\n\u003C\u002Fdiv>\n\n\u003Cdiv align=\"center\">\n\u003Ch3>🚀 τ³-bench is here!\u003C\u002Fh3>\n\u003Cp>From text-only to multimodal, knowledge-aware agent evaluation.\u003Cbr>\nVoice full-duplex · Knowledge retrieval · 75+ task fixes\u003Cbr>\n\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2603.13686\">τ-Voice paper\u003C\u002Fa> · \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2603.04370\">τ-Knowledge paper\u003C\u002Fa> · \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2512.07850\">Task fixes paper\u003C\u002Fa> · \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fsierra-research\u002Ftau2-bench\u002Freleases\u002Ftag\u002Fv1.0.0\">Release notes\u003C\u002Fa>\u003C\u002Fp>\n\u003C\u002Fdiv>\n\n> **How do you say $\\tau^3$-bench?** We just say \"tau three,\" but you do you!\n\n## What's New in $\\tau^3$-bench\n\n- **Knowledge Domain (`banking_knowledge`)** — A knowledge-retrieval-based customer service domain with configurable RAG pipelines, document search, embeddings, and agentic shell-based search. [Learn more →](src\u002Ftau2\u002Fknowledge\u002FREADME.md)\n- **Voice Full-Duplex (Audio Native)** — End-to-end voice evaluation with realtime providers (OpenAI, Gemini, xAI). [Learn more →](src\u002Ftau2\u002Fvoice\u002FREADME.md)\n- **Task Quality (75+ fixes)** — Removed incorrect expected actions, clarified ambiguous instructions, fixed impossible constraints, and added missing fallback behaviors across airline, retail, and banking domains. Based on analysis from [SABER](https:\u002F\u002Farxiv.org\u002Fabs\u002F2512.07850) (Cuadron et al., 2025). [Learn more →](https:\u002F\u002Ftaubench.com\u002Fblog\u002Ftau3-task-fixes.html)\n- **Updated Leaderboard** — Now includes voice and knowledge results. Compare model performance at [taubench.com](https:\u002F\u002Ftaubench.com). [Submit your results →](docs\u002Fleaderboard-submission.md)\n\nSee [CHANGELOG.md](CHANGELOG.md) for the full version history.\n\n> **Backward compatibility note**: If you are evaluating an agent (not training), use the `base` task split to evaluate on the complete task set that matches the original τ-bench structure. This is the default.\n\n> **Upgrading from $\\tau^2$-bench?** Installation now uses `uv` instead of `pip install -e .`, and Python `>=3.12, \u003C3.14` is required (was `>=3.10`). Some internal APIs have been refactored — see [CHANGELOG.md](CHANGELOG.md) for details.\n\n## Overview\n\n$\\tau$-bench is a simulation framework for evaluating customer service agents across multiple domains. It supports text-based half-duplex (turn-based) evaluation and voice full-duplex (simultaneous) evaluation using real-time audio APIs.\n\nEach domain specifies:\n- A **policy** that the agent must follow\n- A set of **tools** that the agent can use\n- A set of **tasks** to evaluate the agent's performance\n- Optionally: a set of **user tools** for the user simulator\n\n**Available domains**: `mock` · `airline` · `retail` · `telecom` · `banking_knowledge`\n\n| Mode | Description |\n|------|-------------|\n| **Text (half-duplex)** | Turn-based chat with tool use |\n| **Voice (full-duplex)** | End-to-end audio via realtime providers (OpenAI, Gemini, xAI) |\n\n## Quick Start\n\n### 1. Install\n\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002Fsierra-research\u002Ftau2-bench\ncd tau2-bench\nuv sync                        # core only (text-mode: airline, retail, telecom, mock)\n```\n\nOptional extras (install what you need):\n\n```bash\nuv sync --extra voice          # + voice\u002Faudio-native features\nuv sync --extra knowledge      # + banking_knowledge domain (retrieval pipeline)\nuv sync --extra gym            # + gymnasium RL interface\nuv sync --extra dev            # + pytest, ruff, pre-commit (required for contributing)\nuv sync --all-extras           # everything\n```\n\nThis requires [uv](https:\u002F\u002Fdocs.astral.sh\u002Fuv\u002Fgetting-started\u002Finstallation\u002F). Voice features also need system dependencies (`brew install portaudio ffmpeg` on macOS). See the [full installation guide](docs\u002Fgetting-started.md) for details.\n\n### 2. Set up API keys\n\n```bash\ncp .env.example .env\n# Edit .env with your API keys (uses LiteLLM — any supported provider works)\n```\n\n### 3. Run an evaluation\n\n```bash\ntau2 run --domain airline --agent-llm gpt-4.1 --user-llm gpt-4.1 \\\n  --num-trials 1 --num-tasks 5\n```\n\nResults are saved to `data\u002Fsimulations\u002F`. Use `tau2 view` to browse them.\n\n> **Tip**: Run `tau2 intro` for an overview of available domains, commands, and examples.\n\n## Documentation\n\n### Getting Started\n\n| Document | Description |\n|----------|-------------|\n| [Getting Started](docs\u002Fgetting-started.md) | Installation, API keys, first run, output structure, configuration |\n| [CLI Reference](docs\u002Fcli-reference.md) | All `tau2` commands and options |\n\n### Core Concepts\n\n| Document | Description |\n|----------|-------------|\n| [Agent Developer Guide](src\u002Ftau2\u002Fagent\u002FREADME.md) | Build and evaluate your own agent |\n| [Domains](src\u002Ftau2\u002Fdomains\u002FREADME.md) | Domain structure, data format, and available domains |\n| [Orchestrator & Communication Modes](src\u002Ftau2\u002Forchestrator\u002FREADME.md) | Half-duplex and full-duplex orchestration |\n\n### Knowledge Retrieval\n\n| Document | Description |\n|----------|-------------|\n| [Knowledge Retrieval](src\u002Ftau2\u002Fknowledge\u002FREADME.md) | Retrieval pipeline configs, embeddings, RAG, and sandbox setup for the `banking_knowledge` domain |\n\n### Voice & Audio\n\n| Document | Description |\n|----------|-------------|\n| [Voice (Full-Duplex)](src\u002Ftau2\u002Fvoice\u002FREADME.md) | Providers, speech complexity, CLI options, and output structure for voice evaluation |\n| [Audio Native Architecture](src\u002Ftau2\u002Fvoice\u002Faudio_native\u002FREADME.md) | Internal architecture for adding or modifying realtime provider adapters |\n\n### RL & Training\n\n| Document | Description |\n|----------|-------------|\n| [Gym Interface](src\u002Ftau2\u002Fgym\u002FREADME.md) | Gymnasium-compatible environment, play mode, train\u002Ftest splits |\n\n### Leaderboard & Experiments\n\n| Document | Description |\n|----------|-------------|\n| [Leaderboard Submission](docs\u002Fleaderboard-submission.md) | How to submit results to [taubench.com](https:\u002F\u002Ftaubench.com) |\n| [Experiments](src\u002Fexperiments\u002FREADME.md) | Experimental features and research code |\n\n### Project\n\n| Document | Description |\n|----------|-------------|\n| [Contributing](CONTRIBUTING.md) | How to contribute to τ-bench |\n| [Changelog](CHANGELOG.md) | Version history and release notes |\n\n## Contributing\n\nWe welcome contributions! Whether you're fixing bugs, adding features, creating domains, or contributing research code, see our [Contributing Guide](CONTRIBUTING.md) for guidelines.\n\n## Citation\n\nIf you use a specific component of $\\tau^3$-bench, please cite the corresponding paper below.\n\n### Knowledge Domain (`banking_knowledge`)\n\n```bibtex\n@article{shi2026tau,\n  title={$\\tau$-Knowledge: Evaluating Conversational Agents over Unstructured Knowledge},\n  author={Shi, Quan and Zytek, Alexandra and Razavi, Pedram and Narasimhan, Karthik and Barres, Victor},\n  journal={arXiv preprint arXiv:2603.04370},\n  year={2026}\n}\n```\n\n### Voice Full-Duplex Benchmark\n\n```bibtex\n\n@misc{ray2026tauvoicebenchmarkingfullduplexvoice,\n      title={$\\tau$-Voice: Benchmarking Full-Duplex Voice Agents on Real-World Domains},\n      author={Soham Ray and Keshav Dhandhania and Victor Barres and Karthik Narasimhan},\n      year={2026},\n      eprint={2603.13686},\n      archivePrefix={arXiv},\n      primaryClass={cs.SD},\n      url={https:\u002F\u002Farxiv.org\u002Fabs\u002F2603.13686},\n}\n```\n\n### Core $\\tau$-Bench\n\n```bibtex\n\n@misc{barres2025tau2,\n      title={$\\tau^2$-Bench: Evaluating Conversational Agents in a Dual-Control Environment}, \n      author={Victor Barres and Honghua Dong and Soham Ray and Xujie Si and Karthik Narasimhan},\n      year={2025},\n      eprint={2506.07982},\n      archivePrefix={arXiv},\n      primaryClass={cs.AI},\n      url={https:\u002F\u002Farxiv.org\u002Fabs\u002F2506.07982}, \n}\n\n@misc{yao2024tau,\n      title={$\\tau$-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains}, \n      author={Shunyu Yao and Noah Shinn and Pedram Razavi and Karthik Narasimhan},\n      year={2024},\n      eprint={2406.12045},\n      archivePrefix={arXiv},\n      primaryClass={cs.AI},\n      url={https:\u002F\u002Farxiv.org\u002Fabs\u002F2406.12045}, \n}\n```\n\n### Task Fixes\n\n```bibtex\n\n@inproceedings{cuadron2026saber,\n      title={{SABER}: Small Actions, Big Errors {\\textemdash} Safeguarding Mutating Steps in {LLM} Agents},\n      author={Alejandro Cuadron and Pengfei Yu and Yang Liu and Arpit Gupta},\n      booktitle={ICLR 2026 Workshop on Memory for LLM-Based Agentic Systems},\n      year={2026},\n      url={https:\u002F\u002Fopenreview.net\u002Fforum?id=En2z9dckgP},\n}\n```\n","# τ-基准：面向真实世界领域的工具-智能体-用户交互基准测试\n\n[![python](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FPython-3.12%2B-blue.svg?style=flat&logo=python&logoColor=white)](https:\u002F\u002Fwww.python.org)\n[![Ruff](https:\u002F\u002Fimg.shields.io\u002Fendpoint?url=https:\u002F\u002Fraw.githubusercontent.com\u002Fastral-sh\u002Fruff\u002Fmain\u002Fassets\u002Fbadge\u002Fv2.json)](https:\u002F\u002Fgithub.com\u002Fastral-sh\u002Fruff)\n[![arXiv](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fcs.AI-arXiv%3A2506.07982-B31B1B.svg?logo=arxiv&logoColor=red)](https:\u002F\u002Farxiv.org\u002Fabs\u002F2506.07982)\n[![blog](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fblog-tau--bench-green)](https:\u002F\u002Fsierra.ai\u002Fblog\u002Fbenchmarking-agents-in-collaborative-real-world-scenarios)\n[![Twitter](https:\u002F\u002Fimg.shields.io\u002Ftwitter\u002Furl\u002Fhttps\u002Ftwitter.com\u002Fsierra.svg?style=social&label=Follow%20%40SierraPlatform)](https:\u002F\u002Fx.com\u002FSierraPlatform\u002Fstatus\u002F1932464265207889974)\n[![LinkedIn](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLinkedIn-0077B5?logo=linkedin&logoColor=white)](https:\u002F\u002Fwww.linkedin.com\u002Fposts\u002Fsierra_last-year-we-introduced-%F0%9D%9C%8F-bench-a-benchmark-activity-7338229693898231809-F8L4?utm_source=share&utm_medium=member_desktop&rcm=ACoAAAdc8goBmhEsiEo1_t_XSJbAnY4_zMfAWcE)\n[![排行榜](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F🏆_Live_Leaderboard-taubench.com-brightgreen?style=flat)](https:\u002F\u002Ftaubench.com)\n\n\u003Cdiv align=\"center\">\n\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fsierra-research_tau2-bench_readme_018677adf914.png\" width=\"95%\" alt=\"轨迹\">\n\u003C\u002Fdiv>\n\n\u003Cdiv align=\"center\">\n\u003Ch3>🚀 τ³-基准来了！\u003C\u002Fh3>\n\u003Cp>从纯文本到多模态、知识感知的智能体评估。\u003Cbr>\n语音全双工 · 知识检索 · 75+项任务修复\u003Cbr>\n\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2603.13686\">τ-语音论文\u003C\u002Fa> · \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2603.04370\">τ-知识论文\u003C\u002Fa> · \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2512.07850\">任务修复论文\u003C\u002Fa> · \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fsierra-research\u002Ftau2-bench\u002Freleases\u002Ftag\u002Fv1.0.0\">发布说明\u003C\u002Fa>\u003C\u002Fp>\n\u003C\u002Fdiv>\n\n> **τ³-基准怎么读？** 我们就叫“塔乌三”，不过你怎么念都行！\n\n## τ³-基准的新特性\n\n- **知识领域（`banking_knowledge`）** — 基于知识检索的客服领域，配备可配置的RAG管道、文档搜索、嵌入以及基于智能体外壳的搜索。[了解更多 →](src\u002Ftau2\u002Fknowledge\u002FREADME.md)\n- **语音全双工（音频原生）** — 使用实时API（OpenAI、Gemini、xAI）进行端到端的语音评估。[了解更多 →](src\u002Ftau2\u002Fvoice\u002FREADME.md)\n- **任务质量（75+项修复）** — 移除了不正确的预期动作，澄清了模糊的指令，修复了不可能的约束，并在航空、零售和银行领域中添加了缺失的回退行为。基于[SABER](https:\u002F\u002Farxiv.org\u002Fabs\u002F2512.07850)（Cuadron等人，2025年）的分析。[了解更多 →](https:\u002F\u002Ftaubench.com\u002Fblog\u002Ftau3-task-fixes.html)\n- **更新后的排行榜** — 现在包括语音和知识评估结果。可在[taubench.com](https:\u002F\u002Ftaubench.com)上比较模型性能。[提交你的结果 →](docs\u002Fleaderboard-submission.md)\n\n完整版本历史请参阅[CHANGELOG.md](CHANGELOG.md)。\n\n> **向后兼容性说明**：如果你只是评估智能体（而非训练），请使用`base`任务划分来评估与原始τ-基准结构匹配的完整任务集。这是默认设置。\n\n> **从τ²-基准升级吗？** 安装现在使用`uv`而不是`pip install -e .`，且需要Python `>=3.12, \u003C3.14`（之前是`>=3.10`）。部分内部API已被重构——详情请参阅[CHANGELOG.md](CHANGELOG.md)。\n\n## 概述\n\nτ-基准是一个用于评估跨多个领域的客服智能体的仿真框架。它支持基于文本的半双工（轮流式）评估，以及使用实时音频API进行的语音全双工（同时进行）评估。\n\n每个领域指定：\n- 智能体必须遵循的**策略**\n- 智能体可以使用的**工具**集合\n- 用于评估智能体表现的**任务**集合\n- 可选：用户模拟器的**用户工具**集合\n\n**可用领域**：`mock` · `airline` · `retail` · `telecom` · `banking_knowledge`\n\n| 模式 | 描述 |\n|------|-------------|\n| **文本（半双工）** | 轮流式聊天并使用工具 |\n| **语音（全双工）** | 通过实时提供商（OpenAI、Gemini、xAI）进行端到端的音频 |\n\n## 快速入门\n\n### 1. 安装\n\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002Fsierra-research\u002Ftau2-bench\ncd tau2-bench\nuv sync                        # 仅核心（文本模式：航空、零售、电信、mock）\n```\n\n可选扩展（按需安装）：\n\n```bash\nuv sync --extra voice          # + 语音\u002F音频原生功能\nuv sync --extra knowledge      # + banking_knowledge领域（检索管道）\nuv sync --extra gym            # + gymnasium RL接口\nuv sync --extra dev            # + pytest、ruff、pre-commit（贡献时必需）\nuv sync --all-extras           # 全部\n```\n\n这需要[uv](https:\u002F\u002Fdocs.astral.sh\u002Fuv\u002Fgetting-started\u002Finstallation\u002F)。语音功能还需要系统依赖（macOS上需`brew install portaudio ffmpeg`）。详细信息请参阅[完整安装指南](docs\u002Fgetting-started.md)。\n\n### 2. 设置API密钥\n\n```bash\ncp .env.example .env\n# 编辑.env文件，填入你的API密钥（使用LiteLLM — 任何支持的提供商均可）\n```\n\n### 3. 运行评估\n\n```bash\ntau2 run --domain airline --agent-llm gpt-4.1 --user-llm gpt-4.1 \\\n  --num-trials 1 --num-tasks 5\n```\n\n结果会保存到`data\u002Fsimulations\u002F`目录下。使用`tau2 view`可以浏览这些结果。\n\n> **提示**：运行`tau2 intro`可获得可用领域、命令和示例的概览。\n\n## 文档\n\n### 入门指南\n\n| 文档 | 描述 |\n|----------|-------------|\n| [入门指南](docs\u002Fgetting-started.md) | 安装、API密钥、首次运行、输出结构、配置 |\n| [CLI参考](docs\u002Fcli-reference.md) | 所有`tau2`命令和选项 |\n\n### 核心概念\n\n| 文档 | 描述 |\n|----------|-------------|\n| [智能体开发者指南](src\u002Ftau2\u002Fagent\u002FREADME.md) | 构建并评估你自己的智能体 |\n| [领域](src\u002Ftau2\u002Fdomains\u002FREADME.md) | 领域结构、数据格式及可用领域 |\n| [编排器与通信模式](src\u002Ftau2\u002Forchestrator\u002FREADME.md) | 半双工和全双工编排 |\n\n### 知识检索\n\n| 文档 | 描述 |\n|----------|-------------|\n| [知识检索](src\u002Ftau2\u002Fknowledge\u002FREADME.md) | 检索管道配置、嵌入、RAG以及`banking_knowledge`领域的沙盒设置 |\n\n### 语音与音频\n\n| 文档 | 描述 |\n|----------|-------------|\n| [语音（全双工）](src\u002Ftau2\u002Fvoice\u002FREADME.md) | 提供商、语音复杂度、CLI选项及语音评估的输出结构 |\n| [音频原生架构](src\u002Ftau2\u002Fvoice\u002Faudio_native\u002FREADME.md) | 用于添加或修改实时提供商适配器的内部架构 |\n\n### RL与训练\n\n| 文档 | 描述 |\n|----------|-------------|\n| [Gym接口](src\u002Ftau2\u002Fgym\u002FREADME.md) | 与Gymnasium兼容的环境、游玩模式、训练\u002F测试划分 |\n\n### 排行榜与实验\n\n| 文档 | 描述 |\n|----------|-------------|\n| [排行榜提交](docs\u002Fleaderboard-submission.md) | 如何将结果提交至 [taubench.com](https:\u002F\u002Ftaubench.com) |\n| [实验](src\u002Fexperiments\u002FREADME.md) | 实验性功能和研究代码 |\n\n### 项目\n\n| 文档 | 描述 |\n|----------|-------------|\n| [贡献指南](CONTRIBUTING.md) | 如何为 τ-bench 做出贡献 |\n| [变更日志](CHANGELOG.md) | 版本历史与发布说明 |\n\n## 贡献\n\n我们欢迎各种形式的贡献！无论您是修复 bug、添加新功能、创建领域，还是贡献研究代码，请参阅我们的 [贡献指南](CONTRIBUTING.md)，以了解相关规范。\n\n## 引用\n\n如果您使用了 $\\tau^3$-bench 的特定组件，请引用下方对应的论文。\n\n### 知识领域 (`banking_knowledge`)\n\n```bibtex\n@article{shi2026tau,\n  title={$\\tau$-Knowledge: Evaluating Conversational Agents over Unstructured Knowledge},\n  author={Shi, Quan and Zytek, Alexandra and Razavi, Pedram and Narasimhan, Karthik and Barres, Victor},\n  journal={arXiv preprint arXiv:2603.04370},\n  year={2026}\n}\n```\n\n### 语音全双工基准测试\n\n```bibtex\n\n@misc{ray2026tauvoicebenchmarkingfullduplexvoice,\n      title={$\\tau$-Voice: Benchmarking Full-Duplex Voice Agents on Real-World Domains},\n      author={Soham Ray and Keshav Dhandhania and Victor Barres and Karthik Narasimhan},\n      year={2026},\n      eprint={2603.13686},\n      archivePrefix={arXiv},\n      primaryClass={cs.SD},\n      url={https:\u002F\u002Farxiv.org\u002Fabs\u002F2603.13686},\n}\n```\n\n### 核心 $\\tau$-Bench\n\n```bibtex\n\n@misc{barres2025tau2,\n      title={$\\tau^2$-Bench: Evaluating Conversational Agents in a Dual-Control Environment}, \n      author={Victor Barres and Honghua Dong and Soham Ray and Xujie Si and Karthik Narasimhan},\n      year={2025},\n      eprint={2506.07982},\n      archivePrefix={arXiv},\n      primaryClass={cs.AI},\n      url={https:\u002F\u002Farxiv.org\u002Fabs\u002F2506.07982}, \n}\n\n@misc{yao2024tau,\n      title={$\\tau$-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains}, \n      author={Shunyu Yao and Noah Shinn and Pedram Razavi and Karthik Narasimhan},\n      year={2024},\n      eprint={2406.12045},\n      archivePrefix={arXiv},\n      primaryClass={cs.AI},\n      url={https:\u002F\u002Farxiv.org\u002Fabs\u002F2406.12045}, \n}\n```\n\n### 任务修复\n\n```bibtex\n\n@inproceedings{cuadron2026saber,\n      title={{SABER}: Small Actions, Big Errors {\\textemdash} Safeguarding Mutating Steps in {LLM} Agents},\n      author={Alejandro Cuadron and Pengfei Yu and Yang Liu and Arpit Gupta},\n      booktitle={ICLR 2026 Workshop on Memory for LLM-Based Agentic Systems},\n      year={2026},\n      url={https:\u002F\u002Fopenreview.net\u002Fforum?id=En2z9dckgP},\n}\n```","# τ²-Bench 快速上手指南\n\nτ²-Bench（含最新 τ³ 扩展）是一个用于评估真实世界领域中“工具 - 智能体 - 用户”交互的仿真框架。它支持文本（半双工）和语音（全双工）两种模式，涵盖航空、零售、电信及银行知识检索等多个领域。\n\n## 1. 环境准备\n\n在开始之前，请确保您的开发环境满足以下要求：\n\n*   **操作系统**：Linux, macOS 或 Windows (WSL2 推荐)\n*   **Python 版本**：必须为 **3.12** 或 **3.13** (`>=3.12, \u003C3.14`)\n*   **包管理器**：必须安装 **[uv](https:\u002F\u002Fdocs.astral.sh\u002Fuv\u002Fgetting-started\u002Finstallation\u002F)** (替代传统的 pip)\n    *   安装命令 (macOS\u002FLinux): `curl -LsSf https:\u002F\u002Fastral.sh\u002Fuv\u002Finstall.sh | sh`\n    *   安装命令 (Windows PowerShell): `powershell -c \"irm https:\u002F\u002Fastral.sh\u002Fuv\u002Finstall.ps1 | iex\"`\n*   **语音功能额外依赖** (如需测试 Voice Full-Duplex)：\n    *   macOS: `brew install portaudio ffmpeg`\n    *   Linux: `sudo apt-get install portaudio19-dev ffmpeg`\n    *   Windows: 需手动安装对应开发库或使用 WSL2\n\n> **注意**：国内开发者若遇到 `uv sync` 下载慢的问题，可尝试配置 `UV_INDEX_URL` 环境变量指向国内镜像源（如清华源或阿里源），例如：`export UV_INDEX_URL=https:\u002F\u002Fpypi.tuna.tsinghua.edu.cn\u002Fsimple`。\n\n## 2. 安装步骤\n\n克隆仓库并根据需求安装依赖。默认安装仅包含核心文本模式（航空、零售、电信、模拟领域）。\n\n```bash\n# 1. 克隆仓库\ngit clone https:\u002F\u002Fgithub.com\u002Fsierra-research\u002Ftau2-bench\ncd tau2-bench\n\n# 2. 基础安装 (仅文本模式)\nuv sync\n\n# 3. 可选：按需安装额外功能\n# 安装语音\u002F音频原生功能 (Voice Full-Duplex)\nuv sync --extra voice\n\n# 安装银行知识检索领域 (Knowledge Domain)\nuv sync --extra knowledge\n\n# 安装 Gymnasium RL 接口\nuv sync --extra gym\n\n# 安装所有功能 (包含开发工具)\nuv sync --all-extras\n```\n\n## 3. 基本使用\n\n### 第一步：配置 API 密钥\n\n框架使用 LiteLLM 作为后端，支持多种大模型提供商。复制示例配置文件并填入您的密钥。\n\n```bash\ncp .env.example .env\n```\n\n编辑 `.env` 文件，填入必要的 API Key（例如 `OPENAI_API_KEY` 等）。\n\n### 第二步：运行评估\n\n使用 `tau2 run` 命令启动评估。以下是一个最简单的示例，在航空领域使用 GPT-4.1 分别作为智能体和用户模拟器，运行 5 个任务：\n\n```bash\ntau2 run --domain airline --agent-llm gpt-4.1 --user-llm gpt-4.1 \\\n  --num-trials 1 --num-tasks 5\n```\n\n*   `--domain`: 选择领域 (`airline`, `retail`, `telecom`, `mock`, `banking_knowledge`)\n*   `--agent-llm`: 被测智能体的模型名称\n*   `--user-llm`: 模拟用户的模型名称\n*   `--num-trials`: 每个任务的重复次数\n*   `--num-tasks`: 运行的任务数量\n\n### 第三步：查看结果\n\n评估结果将自动保存至 `data\u002Fsimulations\u002F` 目录。您可以使用内置命令浏览结果：\n\n```bash\ntau2 view\n```\n\n> **提示**：运行 `tau2 intro` 可查看所有可用领域、命令详解及更多示例。","某金融科技团队正在开发一款能处理复杂银行业务的智能客服 Agent，需要在上线前验证其能否准确调用内部工具并遵守合规政策。\n\n### 没有 tau2-bench 时\n- **测试场景单一**：仅能用简单的文本问答进行测试，无法模拟真实用户在电话中打断、插话的全双工语音交互场景。\n- **知识检索黑盒**：难以评估 Agent 在涉及具体银行条款（如利率政策、转账限额）时，是否能正确通过 RAG 检索文档而非胡乱编造。\n- **规则遵循风险**：缺乏系统化的方法验证 Agent 是否严格遵守“不得向未授权用户透露余额”等关键业务策略，容易埋下合规隐患。\n- **反馈模糊低效**：测试失败后只能看到最终结果错误，无法复现和分析 Agent 调用工具的具体轨迹，导致调试如同大海捞针。\n\n### 使用 tau2-bench 后\n- **全真语音演练**：利用 `voice` 模块接入实时音频 API，直接在模拟用户随时插话的嘈杂环境中，压力测试 Agent 的响应稳定性。\n- **知识能力量化**：通过 `banking_knowledge` 领域任务，精确测量 Agent 检索内部文档的准确率，确保回答有据可依。\n- **策略执行可视**：在 `airline` 或 `retail` 等预设场景中，自动检测 Agent 是否越权操作，将隐性的合规风险转化为显性的评测分数。\n- **轨迹深度诊断**：借助详细的交互轨迹记录，快速定位是工具参数传错还是逻辑判断失误，将模型迭代周期从数天缩短至数小时。\n\ntau2-bench 将原本依赖人工经验的模糊测试，转变为可量化、全覆盖的自动化评估体系，确保智能代理在真实世界中既聪明又守规矩。","https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fsierra-research_tau2-bench_018677ad.png","sierra-research","Sierra Research","https:\u002F\u002Foss.gittoolsai.com\u002Favatars\u002Fsierra-research_9deb1f4b.png","",null,"SierraPlatform","https:\u002F\u002Fsierra.ai","https:\u002F\u002Fgithub.com\u002Fsierra-research",[84,88,92,96,100,104,108],{"name":85,"color":86,"percentage":87},"Python","#3572A5",82.9,{"name":89,"color":90,"percentage":91},"HTML","#e34c26",7.9,{"name":93,"color":94,"percentage":95},"JavaScript","#f1e05a",5.5,{"name":97,"color":98,"percentage":99},"CSS","#663399",3.5,{"name":101,"color":102,"percentage":103},"Makefile","#427819",0.1,{"name":105,"color":106,"percentage":107},"Shell","#89e051",0,{"name":109,"color":110,"percentage":107},"Procfile","#3B2F63",940,239,"2026-04-03T08:19:25","MIT","macOS, Linux, Windows","未说明",{"notes":118,"python":119,"dependencies":120},"1. 安装必须使用 'uv' 工具替代传统的 pip。2. 若启用语音全双工功能，macOS 需通过 brew 安装 portaudio 和 ffmpeg，其他系统需查看完整安装指南配置相应系统依赖。3. 核心功能仅需文本模式，语音和知识库检索功能需通过额外参数 (--extra voice \u002F --extra knowledge) 单独安装。4. 运行前需配置 .env 文件填入 LLM API 密钥。",">=3.12, \u003C3.14",[121,122,123,124,125,126,127,128],"uv (包管理器)","LiteLLM","portaudio (语音功能系统依赖)","ffmpeg (语音功能系统依赖)","pytest (开发依赖)","ruff (开发依赖)","pre-commit (开发依赖)","gymnasium (可选，RL 接口)",[13,54,26,14,15],[131,132,133,134,135],"benchmark","llm","ai","language-model-agent","conversational-agents","2026-03-27T02:49:30.150509","2026-04-06T07:15:03.731156",[139,144,149,154,159,163],{"id":140,"question_zh":141,"answer_zh":142,"source_url":143},10664,"在虚拟环境（.venv）中运行时出现 'FileNotFoundError: tasks.json' 错误怎么办？","这是因为路径推断逻辑在虚拟环境中失效。解决方法有两种：\n1. 使用可编辑模式安装：运行 `pip install -e .` 而不是 `pip install .`，这样会直接使用源代码而非复制到的 site-packages 目录。\n2. 更新版本：维护者已发布新版本修复了此问题，现在支持直接运行或通过环境变量指定数据目录。\n3. 临时方案：手动将 `data\u002F` 文件夹复制到虚拟环境的相应目录中。","https:\u002F\u002Fgithub.com\u002Fsierra-research\u002Ftau2-bench\u002Fissues\u002F7",{"id":145,"question_zh":146,"answer_zh":147,"source_url":148},10665,"运行测试时提示 'ModuleNotFoundError: No module named gymnasium' 如何解决？","这是因为 `gymnasium` 未被列为依赖项。请手动运行以下命令安装：\n`pip install gymnasium`\n维护者已在后续版本中修复了此依赖缺失问题。","https:\u002F\u002Fgithub.com\u002Fsierra-research\u002Ftau2-bench\u002Fissues\u002F87",{"id":150,"question_zh":151,"answer_zh":152,"source_url":153},10666,"Airline 领域的奖励计算基准（reward_basis）为什么只包含 DB 和 COMMUNICATE？","目前 Airline 领域未将 ACTION 和 NL_ASSERTION 纳入最终评分，是因为这些检查依赖 LLM 判断而非程序化断言。团队尚未对 LLM Judge 在这些断言上的表现进行充分校准和评估。一旦研究完成并确认其可靠性，可能会将其集成到最终分数中。","https:\u002F\u002Fgithub.com\u002Fsierra-research\u002Ftau2-bench\u002Fissues\u002F13",{"id":155,"question_zh":156,"answer_zh":157,"source_url":158},10667,"如何为现有领域（如 airline, retail）创建新的编程任务？","所有代码均已开源。你可以参考 `telecom` 领域的实现方式，任务创建和组合代码位于：\n`src\u002Ftau2\u002Fdomains\u002Ftelecom\u002Ftasks`\n虽然代码目前在 telecom 目录下，但其逻辑并不特定于电信领域，可以复用或修改以扩展其他领域（如 airline 或 retail）的任务。","https:\u002F\u002Fgithub.com\u002Fsierra-research\u002Ftau2-bench\u002Fissues\u002F21",{"id":160,"question_zh":161,"answer_zh":162,"source_url":143},10668,"运行命令时复制粘贴失败或报错，可能是什么原因？","README 中的示例命令可能存在格式问题（如换行符后有多余空格），导致直接复制粘贴到 shell 执行失败。建议手动输入或使用以下修正后的命令格式（确保反斜杠后无多余空格）：\n```\ntau2 run \\\n--domain airline \\\n--agent-llm gpt-4.1 \\\n--user-llm gpt-4.1 \\\n--num-trials 1 \\\n--num-tasks 1\n```",{"id":164,"question_zh":165,"answer_zh":166,"source_url":167},10669,"在 solo 模式下运行 gym 环境时报错 'TypeError: missing 1 required positional argument: llm' 怎么办？","这是一个已知 Bug。在 `src\u002Ftau2\u002Fgym\u002Fgym_agent.py` 的 solo 模式逻辑中，实例化 `DummyUser()` 时未传递必需的 `llm` 参数（因为它继承自 `UserSimulator`）。该问题已被维护者确认，请关注后续修复版本。临时解决方法是修改源码，在实例化 `DummyUser()` 时传入一个合法的 LLM 实例。","https:\u002F\u002Fgithub.com\u002Fsierra-research\u002Ftau2-bench\u002Fissues\u002F199",[169,174,179,183],{"id":170,"version":171,"summary_zh":172,"released_at":173},71246,"v1.0.0","# τ-bench 1.0.0 — Voice, Knowledge, Task Quality\n\nThis release transforms τ-bench from a text-only agent evaluation framework into a multimodal, knowledge-aware benchmark. You can now evaluate voice agents in realistic conditions and test agents that must retrieve information from large document corpora.\n\n## 🎙️ Voice Evaluation\n\n**Evaluate any domain in voice mode.** All existing domains (airline, retail, telecom, mock) now support full voice evaluation out of the box.\n\n**7 real-time voice providers** with a provider-agnostic adapter system:\n- **Fully supported:** OpenAI Realtime, xAI Grok Voice, Gemini Live  \n- **Experimental:** Nova Sonic, Qwen, Deepgram (cascaded), LiveKit (cascaded)\n\n**Full-duplex conversations** where user and agent speak simultaneously. The simulated user can yield, interrupt, or wait — testing how agents handle overlapping speech and natural conversational dynamics beyond simple turn-taking.\n\n**Realistic audio conditions** via an audio effects pipeline: background noise, burst sounds (car horns, dog barks), telephony compression, and frame drops. Simulate real call-center conditions rather than clean studio audio.\n\n**Quality assurance** with automatic hallucination detection. The system identifies when the simulated user deviates from task instructions and can re-run affected evaluations.\n\n```bash\n# Evaluate a voice agent on retail tasks\ntau2 run --domain retail \\\n  --audio-native \\\n  --audio-native-provider openai \\\n  --audio-native-model gpt-realtime-1.5 \\\n  --num-tasks 5 \\\n  --audio-taps\n```\n\n## 📚 Knowledge Retrieval + Banking Domain\n\n**New evaluation paradigm:** Agents must now find relevant information in large document corpora before they can act, not just follow a fixed policy.\n\n**New `banking_knowledge` domain:**\n- **97 tasks** spanning account management, credit cards, disputes, transfers\n- **698 policy and procedure documents** — only a few are relevant to any given task\n- Tests both information retrieval and transactional tool use together\n\n**12 configurable retrieval strategies** for apples-to-apples comparison:\n- **Offline:** `no_knowledge`, `full_kb`, `golden_retrieval`, `bm25`, `bm25_grep`, `grep_only`\n- **Embedding-backed:** `openai_embeddings`, `qwen_embeddings` (+ reranker\u002Fgrep variants)  \n- **Agentic:** `terminal_use`, `terminal_use_write` (sandboxed shell access)\n\n```bash\n# Evaluate knowledge retrieval with BM25\ntau2 run --domain banking_knowledge \\\n  --retrieval-config bm25 \\\n  --agent-llm gpt-4.1 \\\n  --user-llm gpt-4.1 \\\n  --num-tasks 5\n```\n\n## 🎯 Task Quality (75+ fixes)\n\n**Airline (27 tasks):** Removed incorrect expected actions, clarified ambiguous instructions, fixed impossible constraints, closed policy loopholes, added missing fallback behaviors.\n\n**Retail (26 tasks):** Removed invalid expected actions (e.g., unsupported PayPal refunds), clarified ambiguous instructions, fixed impossible same-item exchanges, added fallback behaviors.\n\n**Banking (20+ tasks):** Corrected required documents, expected actions, and reward calculations. Cleaned up escaping issues across 155+ policy documents. Ported missing tool validations.\n\n## 🛠️ Developer Experience\n\n**Simpler installation** with `uv` and optional dependency groups:\n```bash\nuv sync                    # core text-mode evaluation  \nuv sync --extra voice      # + voice\u002Faudio-native\nuv sync --extra knowledge  # + banking_knowledge domain\nuv sync --all-extras       # everything\n```\n\n**Richer CLI:**\n- `tau2 intro` — guided introduction to domains and capabilities\n- `tau2 view` — improved simulation viewer  \n- `--timeout` — evaluation time limits\n- Multiple results comparison\n\n**Programmatic API** with three levels of control:\n```python\nfrom tau2.runner import run_simulation          # low-level: run one orchestrator\nfrom tau2.runner import build_text_orchestrator  # mid-level: build from config  \nfrom tau2.runner import run_domain              # high-level: full batch pipeline\n```\n\n**Enhanced evaluation:**\n- LLM-based conversation review and quality checks\n- Hallucination detection for user simulator reliability  \n- Per-task summary analysis and diagnostics\n\n**Comprehensive documentation:**\n- [Getting Started](docs\u002Fgetting-started.md) — installation, setup, first run\n- [CLI Reference](docs\u002Fcli-reference.md) — all commands and options\n- [Knowledge Retrieval](src\u002Ftau2\u002Fknowledge\u002FREADME.md) — retrieval pipeline setup\n- [Audio Native Mode](src\u002Ftau2\u002Fvoice\u002Faudio_native\u002FREADME.md) — voice provider integration\n- Per-module READMEs and developer guides throughout\n\n## 🔧 Breaking Changes\n\n- **Installation method:** Now uses `uv` instead of `pip install -e .`\n- **Python requirement:** Now `>=3.12, \u003C3.14` (was `>=3.10`)\n- Some internal APIs refactored (affects custom agent implementations)\n\n## 📊 By the Numbers\n\n- **172 new source files** in `src\u002Ftau2\u002F`\n- **53 new test files** covering voice providers and knowledge retrieval\n- **7 voice providers** with full-duplex support\n- **97 banking tasks** with **698 documents** in the knowl","2026-03-18T07:14:12",{"id":175,"version":176,"summary_zh":177,"released_at":178},71247,"v0.2.0"," Major new features:\r\n   - Live leaderboard at tau-bench.com\r\n   - Interactive model comparison and performance visualization  \r\n   - Mobile-responsive design\r\n   - Comprehensive submission validation system\"\r\n   ","2025-10-06T16:25:53",{"id":180,"version":181,"summary_zh":79,"released_at":182},71248,"v0.1.3","2025-08-26T23:36:24",{"id":184,"version":185,"summary_zh":79,"released_at":186},71249,"v0.1.2","2025-07-17T21:43:58"]