WebArena Author Shuyan Zhou Joins Meta Superintelligence Labs to Build an AI Browser

Zhou holds a computer science degree from Harbin Institute of Technology and a bachelor's degree in computer science from the University of Hong Kong. After finishing her undergraduate studies in 2018 she entered Carnegie Mellon University's Language Technologies Institute, where she completed a master's degree and then a doctorate, graduating in 2024. Her Google Scholar profile lists 9,569 citations and an h-index of 25.

When she began her PhD in 2020, her research narrowed to a single question: how to let AI enter real digital environments and finish tasks on a user's behalf, deciding which page to open, where to click and what to type, and how to adjust after an error. The question had a personal origin. She had given her mother a Chinese-language version of Alexa, which could handle simple commands such as playing music or reporting the weather but could not book flights. Her mother would call Zhou and ask for step-by-step instructions. Zhou set out the idea in her PhD application materials in 2020, and four years later her dissertation, "Using AI Agents to Solve Real-World Tasks," gave a partial answer.

Her best-known product is WebArena, built during her doctoral work as a realistic web environment for autonomous agents, with real interaction and task execution. Before WebArena, many browser agents were tested on simplified pages or static datasets, which made experiments easy but poorly reflected continuous operations, page changes and user permissions on live sites. Zhou and her team used open-source software to rebuild e-commerce, forum, code collaboration and content management sites into independently deployable, reusable interactive environments. They also designed 241 task templates that generate 812 test tasks, checking whether an AI can follow natural-language instructions to search, click, type, navigate between pages and edit content.

VisualWebArena extended the framework from text to images, icons and page layout, adding 910 web tasks that require visual understanding. Zhou then worked with the University of Hong Kong's XLANG team and other institutions on OSWorld, which widened evaluation from the browser to the whole computer system. OSWorld contains 369 computer operation tasks spanning browsers, file systems, office software and coordination across multiple applications. It was first released in April 2024 and was later accepted to the NeurIPS 2024 Datasets and Benchmarks track. At publication, humans completed 72.36 percent of the test tasks, while the best-performing model completed 12.24 percent.

The open problems include visual grounding, long-horizon planning, dynamic environments, error recovery and security permissions, all of which are likely to surface in an AI browser.

This is Zhou's second stint at Meta. In 2024 she joined Meta GenAI, working in a team led by Ruslan Salakhutdinov, who also advised Yang Zhilin at Carnegie Mellon, on Llama general computer-operating agents. She left Meta last year to teach at Duke University and stayed until early this year.

Meta's Superintelligence Labs, or MSL, was founded in June 2025 as the company's consolidated AI organization. It brought under one structure the TBD Lab responsible for next-generation foundation models, the long-term research group FAIR, product and applied research teams, and MSL Infra, which supplies compute and engineering support. The stated goal is "personal superintelligence." In Mark Zuckerberg's account, superintelligence should not replace people but should understand what users need and deliver value through Meta AI, WhatsApp, Instagram and smart glasses.

Meta launched a high-profile recruiting drive last year, offering top pay to senior researchers. Among the managers and researchers MSL has announced publicly, close to half are of Chinese descent. They include MSL head Alexandr Wang, who founded Scale AI at 19 and whose company later drew a $14.3 billion investment from Meta; chief scientist Shengjia Zhao, a Tsinghua graduate who worked on ChatGPT, GPT-4 and reasoning models at OpenAI; and chief researcher Yang Song, a representative diffusion-model researcher who led OpenAI's strategic exploration team and now helps design the lab's research direction alongside Zhao. Shuchao Bi, a Zhejiang University mathematics graduate who was an engineering lead at Google and oversaw multimodal post-training at OpenAI, works on reinforcement learning, model post-training and agents. Dawn Song, a Tsinghua undergraduate alumnus and UC Berkeley professor who studies AI safety and computer security, works on AI safety and trustworthy agents.

After the reorganization and hiring, MSL has begun releasing products. In April this year Meta released Muse Spark, the lab's first model, a multimodal reasoning model designed for personal agent scenarios. In September it launched Muse, a personal agent powered by Muse Spark, which users can ask to handle email, fill in forms, shop or draw up long-term plans, with future integration into AI smart glasses. Zhou's AI browser work overlaps with that product line.