Shunyu Yao

Shunyu Yao (姚顺雨) is an AI researcher whose collaborative work includes ReAct, which interleaves reasoning with actions, and Tree of Thoughts, which explores alternative reasoning paths. His work also addresses how agents use software and how to evaluate their reliability. In April 2026, Tencent identified him as its Chief AI Scientist.[3][5][6][8][9][12]

Yao’s personal website, last updated in February 2025, identified him as an OpenAI researcher and included Computer-Using Agent (CUA) and Deep Research among his selected work. He described CUA as an interface for interacting with the digital world and Deep Research as an agent that navigates and reasons over the Internet. He also appeared alongside Harrison Chase in Latent Space’s “Language Agents: From Reasoning to Acting”.[3][2]

In December 2025, Tencent promoted Yao to chief AI scientist under the CEO’s office. The South China Morning Post reported that he was a recent recruit and former OpenAI researcher, reporting directly to Tencent president Martin Lau and tasked with leading new AI infrastructure efforts, including large-language-model development.[13]

At the April 24, 2026 Hy3 preview release, Yao described expanding Tencent’s pre-training and reinforcement-learning efforts while using preview feedback to improve real-world applicability. He emphasized connecting model development to products: “Through deep co-design with Tencent’s products, we will continue to enhance real-world performance while uncovering differentiated capabilities of the model.”[12]

In his April 10, 2025 essay “The Second Half”, Yao argued that AI research should shift emphasis from solving benchmark problems to defining useful problems: “In this new era, evaluation becomes more important than training.” He called the gap between impressive AI capabilities and practical usefulness the utility problem, which he regarded as AI’s most important problem.[11]

Yao challenged two familiar evaluation assumptions. First, an agent should not always receive one complete task, work alone and return a final answer: customer service often requires continued interaction with a person. Second, evaluating every task independently misses the value of accumulated experience. He contrasted a software engineer becoming familiar with a repository over time with an agent solving issues without gaining that familiarity, arguing for evaluations that make long-term memory useful.[11]

These arguments followed several years of Yao’s work on language-agent environments, reasoning methods and benchmarks. He earned his Princeton computer-science PhD in 2024, advised by Karthik Narasimhan; his thesis was titled Language Agents: From Next-Token Prediction to Digital Automation.[4][3]

Looking back in his 2025 essay, Yao recalled applying supervised fine-tuning and reinforcement learning to GPT-2 for text-based games in 2019. He described slow learning within individual games and poor transfer to new ones. His insight, as he recalled it, was that agents could choose to reason about a situation as well as take actions in it. He summarized this view as “language generalizes through reasoning in agents.”[11]

In 2022, Yao coauthored WebShop with Howard Chen, John Yang and Narasimhan. The simulated shopping environment contained 1.18 million real-world products and 12,087 crowdsourced instructions. An agent had to interpret a product requirement, navigate pages, search, customize an item and complete a simulated purchase—testing language understanding through a sequence of decisions rather than a single answer.[10]

Yao coauthored ReAct, first submitted in October 2022, with Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Narasimhan and Yuan Cao. Its central mechanism interleaves reasoning and acting: reasoning traces help an agent update plans and handle exceptions, while actions gather information from external sources or environments. For question answering and fact verification, the team used a Wikipedia API to ground the model’s reasoning in retrieved information.[5]

With Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao and Narasimhan, Yao coauthored Tree of Thoughts, submitted in May 2023. It treats coherent pieces of text as intermediate reasoning steps, allowing a model to evaluate multiple paths, look ahead and backtrack. In the paper’s Game of 24 experiment, the authors reported 74% success, compared with 4% for GPT-4 using chain-of-thought prompting. Yao discussed the work in Latent Space’s NeurIPS 2023 recap, including a poster-session interview.[6][1]

Yao and Theodore R. Sumers contributed equally to Cognitive Architectures for Language Agents (CoALA), coauthored with Narasimhan and Griffiths; their author order was decided by coin flip. Submitted in September 2023, the framework draws on cognitive science and symbolic AI to organize agents around modular memory, actions that interact with internal memory or external environments, and a decision process for choosing those actions.[7]

In 2024, Yao coauthored SWE-agent with John Yang, Carlos E. Jimenez, Alexander Wettig, Kilian Lieret, Narasimhan and Ofir Press. The team investigated agents as software users with their own interface needs. SWE-agent’s custom agent-computer interface supports creating and editing files, navigating repositories and executing tests, making interface design part of the system’s approach to automated software engineering.[8]

Yao, Noah Shinn, Pedram Razavi and Narasimhan introduced τ-bench in June 2024. It evaluates agents in dynamic conversations with language-model-simulated users, using domain-specific API tools and policy guidelines. Success is checked by comparing the final database state with an annotated goal; the pass^k metric measures reliability across repeated trials. In those experiments, the authors reported that even then-leading function-calling agents such as GPT-4o succeeded on fewer than half the tasks, with retail pass^8 below 25%.[9]