I am 池裕涵, namely Yuhan Chi (pronounced /juːˈhɑːn tʃiː/) in English. I am an undergraduate at the School of Mathematical Sciences, Fudan University.

I am interested in distilling complex engineering problems into elegant mathematical models. Besides, I am always thinking about how to make decisions under uncertainty.

I am also deeply concerned about socioeconomic issues and hope to leave the world a better place. (In fact, I’ve already started working towards this.)

Research Interests

Optimization, control, and reinforcement learning for sequential decision-making.

Mathematical modeling for real-world problems.

Experience

Visiting Student, Mechanical Systems Control Lab, UC Berkeley — Jul – Aug 2026
Engaged in frontier robotics research workflows.

token-verification-mirage

Solo-author · ICML 2026 AI4Math Workshop Poster

Controlled evaluation of token-level verification signals for LLM math reasoning. Protocol choices like global pooling and direction-agnostic AUROC can shift reported performance by up to 0.18. Shallow token statistics are useful diagnostics, not stable standalone verifiers.

Paper · Code · Workshop

code-not-text

Solo measurement study · 31,040 runs across math, science, coding

Tests cheap hand-crafted features from reasoning traces across domains. Strong math signal (AoA 0.958), partial science signal (0.799), weak coding transfer (0.434). Same features track convergence in math but not executable correctness in code.

Code · Demo · Technical note

TinyLoRA-GRPO-Coder

Solo project · RL post-training for code generation

Small-parameter adaptation and GRPO training pipeline for competitive-programming code generation on Qwen2.5-Coder-3B. Uses compile-and-run rewards rather than static heuristics.

Talks

From Bandits to PPO: RL Prerequisite Seminar — Fudan University, Apr 2026 (PDF)</div>

Publications

Token Verification Mirage
ICML 2026 Workshop on AI for Math (AI4Math), 2026