Shunyu Yao

Research

Yao's research concerns language agents, systems in which language models reason and act in external environments. ReAct (2022), developed while Yao was a Student Researcher on Google's Brain Team with joint support from the Princeton NLP Group, has a model interleave reasoning traces with task actions. Tree of Thoughts (2023), which lists Yao with a Princeton University affiliation, lets a model explore and evaluate several reasoning paths before committing to one.²³⁵

With Theodore Sumers, Karthik Narasimhan, and Thomas Griffiths, Yao co-wrote "Cognitive Architectures for Language Agents" (2023), a framework that organizes language agents around memory, action space, and decision-making. Yao's PhD thesis, listed on Yao's homepage, is titled "Language Agents: From Next-Token Prediction to Digital Automation".⁶¹

Benchmarks and environments

Yao has built several environments for testing agents. WebShop (2022) is a simulated online store in which agents follow natural-language shopping instructions, and Yao co-authored SWE-bench (2023), which asks language models to resolve real GitHub issues. τ-bench (2024) emulates conversations between a simulated user and an agent that must use domain-specific API tools while following policy guidelines.⁷⁸⁴⁹

Products and projects

Yao's homepage lists two OpenAI agent products among Yao's selected work: the Computer-Using Agent (CUA), described as a universal interface for interacting with the digital world, and Deep Research, an agent that navigates and reasons over the internet to carry out research.¹

Work and contributions

• τ-bench Open-source code

  First author, with Noah Shinn, Pedram Razavi and Karthik Narasimhan, of "τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains" (2024), which emulates conversations between a language-model-simulated user and an agent with domain-specific API tools and policy guidelines. The sierra-research/tau-bench repository is MIT-licensed.⁴⁹¹

• ReAct Published research

  First author of "ReAct: Synergizing Reasoning and Acting in Language Models" (2022), which has language models generate reasoning traces and task-specific actions in an interleaved manner.²³

Timeline

1. Jul 4, 2022Released WebShop, a benchmark of web shopping tasks for grounded language agents, as first author.⁷
2. Oct 6, 2022Released ReAct as first author, combining reasoning traces and actions in language models.²
3. Nov 8, 2022Co-wrote the Google Research blog post introducing ReAct as a Student Researcher on the Brain Team.³
4. May 17, 2023Released "Tree of Thoughts", on deliberate problem solving with language models, as first author.⁵
5. Sep 5, 2023Co-authored "Cognitive Architectures for Language Agents".⁶
6. Oct 10, 2023Co-authored SWE-bench, a benchmark of real-world GitHub issues for language models.⁸
7. Jun 17, 2024Released τ-bench, a benchmark for tool-agent-user interaction, as first author.⁴

Documented roles

The sources do not confirm a current role; these are the roles they document.

• Researcher, OpenAI¹
• Student Researcher, Google Research (Brain Team)³

Companies in their history

• OpenAI 63 people in the wiki
• Google 121 people in the wiki

Related people

• Collin BurnsConnected to OpenAI
• Mark ChenConnected to OpenAI
• Jason WeiConnected to OpenAI
• Diogo AlmeidaConnected to OpenAI
• Hyung Won ChungConnected to OpenAI
• Thomas ScialomLanguage models · Agents
• Gabe PereyraAgents · Evaluation and benchmarks
• Yoav ShohamLanguage models · Agents
• Eric Michael SmithEvaluation and benchmarks · Language models
• Emmanuel DupouxLanguage models · Evaluation and benchmarks