Shunyu Yao

Shunyu Yao helped define the modern language-model agent through ReAct and Tree of Thoughts, then tested whether such systems can complete long, rule-bound tasks. Now Tencent’s chief scientist for AI, he is directing work on models and benchmarks built for sustained agency.

Shunyu Yao is Tencent’s chief scientist for AI, after research at Princeton and OpenAI. His first-author papers ReAct and Tree of Thoughts supplied two foundational patterns for language-model agency: interleaving reasoning with external actions, and exploring multiple reasoning paths before committing to an answer. In 2025, he extended that agenda through τ-bench, an ICLR paper evaluating agents in realistic retail and airline settings with tools, policies and simulated users. The study found that then-leading agents completed fewer than half of tasks, with pass^8 below 25% in retail, exposing inconsistency across repeated attempts. Yao was also credited as a core research contributor to Deep Research. After joining Tencent in December 2025, he became last author of CL-bench, a 2026 benchmark for continual learning across 500 contexts, 1,899 tasks and 31,607 expert-written rubrics. The paper reported an average task-solve rate of 17.2% across ten frontier models; a follow-up, CL-bench Life, reported 13.8% average and 19.3% for the best tested system. As chief scientist, Yao has publicly presented Tencent’s Hy3 model programme. Its specifications and benchmark comparisons remain company-reported. Across institutions, his contribution is consistent: define an agentic method, then design tests difficult enough to reveal its limits.