CV
View PDF version
Profile
- 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 (Fall 2025), 4.00/4.00 (Spring 2026); Cumulative GPA: 3.98/4.00; Cumulative Rank: 3/213 in cohort.
- Research interests: mathematical interpretability for AI trust, verification/evaluation of LLM reasoning, optimization, control, and reinforcement learning for sequential decision-making.
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 (Fall 2025), 4.00/4.00 (Spring 2026) (Cumulative: 3.98/4.00)
- Cumulative Rank: 3/213 in cohort
Research Interests
- Mathematical Interpretability and AI Trust: grounding claims about model reliability in inspectable mathematical and statistical evidence
- Reasoning Verification and Evaluation: studying when trace-level signals are diagnostic, misleading, protocol-dependent, or domain-specific
- Optimization, Control, and Sequential Decision-Making: using mathematical tools to reason about uncertainty, feedback, constraints, and interaction with environments
Selected Open-Source Projects
- token-verification-mirage — Paper PDF · AI4Math Workshop
- Solo-author project; led the full pipeline from dataset selection and model deployment to trace generation, evaluation design, analysis, figures, related-work organization, and writing
- Poster, ICML 2026 Workshop on AI for Math (AI4Math)
- Audits how evaluation protocol choices such as pooling, in-sample scoring, and direction-agnostic AUROC can change apparent verification performance
- Shows that shallow token statistics can be useful diagnostics, but should not be treated as stable standalone verifiers without stronger controls
- code-not-text — Interactive Demo
- Solo project; cross-domain measurement study of cheap CoT-surface and token-trajectory features on DeepSeek-R1-0528-Qwen3-8B across 31,040 math, science, and coding runs
- Reports strong math signal (AoA 0.958, AUROC@100% 0.982), partial science signal (AoA 0.799), and weak coding transfer (AoA 0.434)
- Tests coding-side robustness through an 83-scalar feature sweep, grouped ablations, a CoT-only judge, nonlinear MLPs, SSL pre-training, semantic-knot annotation, and token-level de-knotting
- Frames the result as measurement non-invariance: the same feature family tracks convergence-like behavior in math but not executable correctness in coding
- TinyLoRA-GRPO-Coder
- Solo project; independent reimplementation and adaptation inspired by Learning to Reason in 13 Parameters
- Moved a small-parameter adaptation and GRPO-style training pipeline toward verifiable competitive-programming code generation on Qwen2.5-Coder-3B
- Uses compile-and-run rewards rather than static heuristics
- Built as a full research-system exercise: data processing, training, multi-GPU setup, reward design, evaluation, and validation
- microgpt.cpp
- Solo project; minimal GPT implementation from first principles in C++, built to understand model internals without relying on high-level frameworks
Academic Service
Skills
- Programming: Python, C++, C, Java, Node.js, Lean 4
- ML/AI: PyTorch, ML experimentation, LLM evaluation, reinforcement-learning basics
- Tools: Git, GitHub, Linux, LaTeX, Markdown, VS Code
- Other: Mathematical proof writing, technical writing, self-directed learning, competitive programming
- Maintains github-unflag-playbook-cn, a Chinese playbook documenting GitHub account flagging, recovery cases, and appeal workflows
- Contributor to FDUGuideBook/nav-site, a student-built navigation site for the Fudan community
- Contributor to FDU-Sharing, contributing course materials, documentation, and small features for an open, mutual-aid course-material sharing project
- Personal website: chi-shan0707.github.io
- GitHub: Chi-Shan0707
- Actively self-studying mathematics, algorithms, and AI beyond formal coursework, with ongoing work documented in public repositories
