Coding agents and agentic RL
Verified SWE tasks (LegoFlow), SFT recipes (SWE-Lego), terminal-agent trajectories (Terminal-Lego), and RL inside Claude Code, OpenHands and OpenCode (Lego-RL, up to 70.4% on SWE-bench Verified).
Research
Chaofan Tao (陶超凡) is an AI researcher at LegoX and Huawei Foundation Model Lab. His research spans three connected areas: training open LLM coding agents end to end, making LLMs cheaper to train and serve, and improving how multimodal models reason and act. He received his PhD from The University of Hong Kong, and his work on generative model quantization received an ACL 2022 Outstanding Paper Award.
Efficient LLM training and inference
Vocabulary scaling laws (NeurIPS 2024), quantization of generative models (ACL 2022 Outstanding Paper), KV cache compression, and fast test-time scaling.
Multimodal reasoning and agents
RL and long-CoT post-training for reasoning VLMs, physical reasoning (PhyX), multimodal browsing agents, and code generation for robots (RoboCodeX).
All 25 publications with English and Chinese summaries are on the homepage. Machine-readable summaries: llms.txt and llms-full.txt.