Taofeng Xue
Staff Researcher at Meituan
Taofeng Xue is a Staff Researcher at Meituan. He received his master’s degree from the Institute of Software, Chinese Academy of Sciences and joined Meituan as a top-tier talent under the “Beidou” Program. At the forefront of large models and autonomous agents, he led— as first author — the industry-leading **EvoCUA (Evolving Computer Use Agents)** project. In January 2026, EvoCUA achieved a 56.7% success rate and ranked No.1 on the OSWorld open-source leaderboard. Following the full open-source release, the model quickly surpassed 12K+ downloads, and its technical report topped the Hugging Face Daily Paper rankings. Before focusing on foundation model innovation, he accumulated over five years of hands-on experience in large-scale search, advertising, and recommendation systems. He previously served as the Ranking Lead for community search at Meituan Dianping. Combining cutting-edge research vision with strong engineering execution, he has led teams to win the championship in multimodal reasoning at ICCV 2025 and secure global runner-up in multimodal recommendation at RecSys Challenge 2024. He has also published multiple papers at top-tier conferences including ICCV, RecSys, SIGIR, and CIKM. In addition, he is a well-known technical blogger under the name “Mr. Mushroom,” with over 50,000 professional followers, continuously contributing high-value AI insights to the open-source community.
Topic
Towards Digital Life: Core Technologies and Self-Evolution Practices of the Computer Use Agent (EvoCUA)
Enabling AI to perceive screens visually, operate keyboard and mouse, and autonomously complete long-horizon, cross-application computer tasks (Computer Use Agent) is a key milestone in the evolution of large models toward “digital life.” However, the industry currently faces significant deployment challenges, including low-quality data synthesis, lack of interaction feedback, and difficulty in long-horizon credit assignment. This talk will provide an in-depth breakdown of the EvoCUA project, which I led and successfully open-sourced, sharing our experience on both algorithmic and infrastructure fronts: 1. **Verifiable Data Synthesis**: Built an “generate-and-verify” agentic engine to scale high-quality tasks across full scenarios. 2. **Hundred-Thousand-Level Sandbox Infrastructure**: Overcame infrastructure bottlenecks to spin up tens of thousands of sandboxes within minutes, supporting high-concurrency interaction exploration. 3. **Experience Evolution Learning**: Integrated cold-start structured reasoning, RFT dynamic denoising, and RL key-point reflection, enabling the model to truly “evolve from failures.” EvoCUA was fully open-sourced in January 2026, achieving a 56.7% success rate on the OSWorld benchmark, significantly advancing the open-source SOTA. The model downloads quickly surpassed 12K+, and the technical report topped the HuggingFace Daily Paper leaderboard (2026.01.23). This presentation will dissect the full pipeline from infrastructure to algorithmic paradigms, providing high-value practical insights for deploying multimodal large model agents in real-world applications.
