Yuhan Chi 池裕涵
Undergraduate student
School of Mathematical Sciences, Fudan University
I am Yuhan Chi (池裕涵), pronounced /juːˈhɑːn tʃiː/, an undergraduate at the School of Mathematical Sciences, Fudan University, studying Information & Computing Science and Artificial Intelligence.
I enjoy distilling complex engineering problems into elegant mathematical models and thinking about how to make decisions under uncertainty. My current research focuses on the intersection of reinforcement learning and generative models.
I also care deeply about socioeconomic issues and contribute to open-source projects for learning, sharing, and community life.
Feel free to reach out! I’m always happy to connect over coffee, sports, or a shared interest.
Research Interests
Optimization, control, and reinforcement learning for sequential decision-making and generative models; mathematical modeling for real-world problems.
Publications
- Calibrating What Behavior Can Identify: Conformal Reward Sets for Inverse Reinforcement Learning NeurIPS 2026 Workshop on Physical Understanding for Decision-Making (PUDM), 2026 Poster Adapts conformal inverse optimization to IRL with a coordinate-invariant, policy-value-based score, characterizes when behavioral compatibility implies true-reward coverage, and proves when robust planning over the resulting set cannot certify improvement over a reference policy. PaperCodeProject
- Token-Level Verification under Controlled Evaluation ICML 2026 Workshop on AI for Math (AI4Math), 2026 Poster Shows that evaluation-protocol choices alone shift the apparent verification AUROC of LLM math traces by up to 0.18; under controlled scoring, shallow token statistics are diagnostics rather than reliable standalone verifiers. PaperCode
Education & Experience
-
Sep 2025 – present
B.S. in Information & Computing Science and Artificial Intelligence (double degree) Fudan University, School of Mathematical SciencesGPA 3.98 / 4.00 · Rank 3 / 228
-
Jul – Aug 2026
Visiting Student Mechanical Systems Control Lab, UC BerkeleyStudied robot hardware interfaces and explored dexterous-hand grasping in simulation.
Selected Projects
-
code-not-text
Can cheap reasoning-trace features predict correctness across math, science, and code? Selecting among 64 candidates with them changes pass@1 by +10.0 pp in math, +8.0 pp in science, and −0.6 pp in code.
-
TinyLoRA-GRPO-Coder
Can a tiny shared parameterization learn code generation from executable feedback? Trains 32 shared scalars of Qwen2.5-Coder-3B with GRPO, rewarded by real compile-and-run outcomes.
Talks, Service & Honors
- Talk From Bandits to PPO: an RL prerequisite seminar, Fudan University, Apr 2026
- Service Reviewer, ICML 2026 Workshop on AI for Math (AI4Math)
- Service Founder, Fudan Open Source Initiative
- Honor First Prize, National Olympiad in Informatics in Provinces (NOIP), Shanghai, 2023