Xin Li

I am a Ph.D. student at Nanyang Technological University (NTU), advised by Prof. Chau Yuen. I work on making large language models (LLMs) reason reliably: benchmarks that measure whether they do, post-training that teaches them to, and agentic systems that test whether coordination actually helps. Most of it lands in domains where an answer can be checked—formal verification, mathematics, and code.

• LLM Evaluation Benchmarks and evaluation protocols built to be audited rather than trusted: per-problem provenance, verifier-checkable answers, and explicit limits on what a score supports—including WirelessMathBench-XL, DafnyComp, WritingPreferenceBench, and RobustMAD.
• Post-training Reinforcement learning, on-policy distillation, and verifier-guided selection, all aimed at reasoning where the reward can be checked rather than guessed—including Re:Form, WirelessMathLM, TLVC, ListOPD, and DN-MOPD.
• Agentic Systems Protocols, memory, and measurement for systems where several models interact: what they send each other, what they keep, and whether the interaction helps or hurts—including LACP and DebateLedger.

Before my Ph.D. I worked on robot perception—visual-inertial odometry at MEGVII, RGB-D + IMU indoor mapping at Microsoft Research Asia, and multimodal localization at Gausium Robotics, where I led a five-engineer team and shipped to a fleet of 1,000+ commercial cleaning robots. I received my M.E. from Peking University and my B.E. from Northeastern University, China.