CV
View PDF version
Profile
- First-year undergraduate at Fudan University (School of Mathematical Sciences), pursuing a double degree in Information and Computing Science and Artificial Intelligence.
- GPA: 3.96/4.00 (most recent semester); ranked 10th (top 5%) in cohort.
- Research focus: mathematical foundations of intelligence, entrustable sequential decision-making, and reinforcement learning with world models.
Education
- Fudan University, Shanghai, China
- B.S. Candidate in Information and Computing Science & Artificial Intelligence (Double Degree), School of Mathematical Sciences
- Sep. 2025 – Present
- GPA: 3.96/4.00 (most recent semester)
- Rank: 10th (top 5%) in cohort
Research Interests
- Mathematical Foundations of Intelligence: grounding AI trust in formal mathematical terms rather than empirical benchmarks
- Entrustable Sequential Decision-Making: building agents that are mathematically verifiable and know when to act, defer, or bear consequences
- Reinforcement Learning & World Models: RL as a language for trust, uncertainty, and structured interaction with environments
Selected Open-Source Projects
- code-not-text — Interactive Demo
- Paper-level research on cross-domain limits of hand-crafted CoT-surface features for predicting solution correctness from reasoning traces
- Evaluated on DeepSeek-R1-0528-Qwen3-8B across math, science, and coding; strong in math (AoA 0.958), below-chance in coding (AoA 0.434)
- Five independent checks (probes, feature sweep, MLPs, SSL, de-knotting) all converge to the same coding ceiling, ruling out capacity, engineering, scarcity, and noise — code correctness is not in the text
- Argued that CoT traces are not entrustable correctness instruments: they signal reasoning failure in math but only text fluency in coding, making execution feedback essential for verifiable agents
- Interactive demo at chi-shan0707.github.io/code-not-text/demo
- TinyLoRA-GRPO-Coder
- Independent reimplementation and adaptation of TinyLoRA + GRPO from Learning to Reason in 13 Parameters
- Migrated the method from math reasoning to verifiable competitive-programming code generation on Qwen2.5-Coder-3B
- Trained with a minimal number of shared trainable parameters; replaced static heuristics with real compile-and-run rewards
- Built the full pipeline end to end: data processing, training, multi-GPU setup, reward design, evaluation, and validation
- microgpt.cpp
- Minimal GPT implementation from first principles in C++, built to understand model internals without relying on high-level frameworks
Skills
- Programming: Python, C++, C, Java, Node.js, Lean 4
- ML/AI: PyTorch, deep learning, LLM fine-tuning, reinforcement learning, computer vision
- Tools: Git, GitHub, Linux, LaTeX, Markdown, VS Code
- Other: Mathematical proof writing, technical writing, self-directed learning, competitive programming
