Profile
I work near the boundary between mathematical structure and modern learning systems: how models generalize, how training dynamics can be read, and how reasoning traces can be attributed or interpreted. My current record is coursework- and project-shaped rather than publication-shaped, but it already has the ingredients I want to keep: clean questions, reproducible baselines, diagnostic experiments, and a habit of explaining why a result matters.
Education
Peking University
School of Mathematical Sciences
Sep 2024 - Present
GPA 3.79 / 4.00. Recent semester GPAs: 3.803, 3.810, 3.763. Core coursework includes Mathematical Analysis I-III, Geometry, Advanced Algebra I-II, Probability Theory, Mathematical Statistics, Data Structures and Algorithms, Introduction to Mathematical Machine Learning, Ordinary Differential Equations, Deep Learning Theory, Multi-Agent Foundations, and Applied Stochastic Processes.
Weifang Beichen High School
Early admission to the PKU Mathematics Talent Program
Sep 2022 - Aug 2024
Honors
Lingjun Linghang University Scholarship, 2025
National College Student Mathematics Competition, First Prize, 2025
China Undergraduate Mathematical Contest in Modeling, Second Prize, 2025
Chinese Mathematical Olympiad, Gold Medal, 2023
Southeast Mathematical Olympiad, Gold Medal, 2023
Chinese High School Mathematics League, First Prize, 2022 and 2023
Research Interests
Learning Dynamics and Scaling
Model Interpretability and Mathematical Reasoning
Algorithmic Generalization
Math Text Attribution
Quantitative Sequence Modeling
Agent Systems and RAG
Selected Coursework and Projects
Final project for Mathematical Introduction to Machine Learning. Projected LR-Drop Residuals studies whether a source cosine loss curve can identify a schedule-response residual that transfers to WSD-family curves without fitting target WSD losses.
- Treated MPL as a frozen baseline and framed the remaining error as an identification problem rather than a generic residual-copying problem.
- Projected out local MPL-LD tangent nuisance directions before estimating a single source-only response amplitude.
- Used schedule-derived response and locality features to keep target losses outside the deployable prediction pipeline.
Causal Transformer for High-Frequency Return Prediction
Project Notes
Lingjun Quant Challenge project. Built a causal Transformer for A-share high-frequency microstructure data with 500 stocks, 239 intraday minutes, and 384 features, predicting ten-minute-ahead returns with strict time-split validation and leakage-aware preprocessing.
- Modeled single-stock single-day minute sequences as tokens with causal masks, intraday time embeddings, and stock identity embeddings.
- Implemented fixed train-date normalization and compared causal test-time update / intraday blended variants.
- Built checkpoint, prediction, residual, time-profile, and parameter diagnostics to understand where the sequence model was actually learning and where it was only fitting noise.
Midterm project for Mathematical Modeling. Built a source-attribution system for mathematical solutions from humans and major LLMs including DeepSeek, GLM, Kimi, and Qwen, combining feature-based baselines with neural text classification.
- Designed zero-shot cross-language transfer experiments and observed feature collapse in traditional statistical features.
- Built an automated data collection, feature engineering, training, and confusion-matrix evaluation pipeline.
- Used attribution as a small-scale probe for how mathematical language, model family, and dataset construction shape apparent reasoning signals.
MultiagentFinal: Intent-Grounded Cooperative Sokoban
Project Notes
Final project for Multi-Agent Foundations. Built StrictCoop-Sokoban and studied intent-grounded recurrent communication for partially observable cooperative multi-agent reinforcement learning.
- Removed action aliasing in cooperative Sokoban by enforcing strict push semantics, planner-verified level pools, local observations, and a disjoint hard evaluation split.
- Designed IGRC-MAPPO with low-dimensional broadcast intent messages, future-box auxiliary grounding, and DRC-style ConvLSTM memory under centralized training and decentralized execution.
- Improved hard-v2 performance from MAPPO's 0.763 ± 0.012 pass@8 to 0.949 ± 0.015 pass@8 across three seeds, with ablations separating communication, grounding, memory, capacity, and masks.
Final project for Selected Topics in Deep Learning Theory. Built a reproducible PyTorch harness for grokking on small algorithmic datasets, covering modular arithmetic, polynomial modular operations, S5 group tasks, and K-ary modular summation with Transformer, MLP, LSTM, and GRU comparisons.
Teaching-assistant project design for Data Structures and Algorithms B. Developed and maintained a Flask-based Avalon agent battle platform where submitted Python agents compete under partial observability, communication constraints, referee-controlled game phases, automatic matching, ELO ranking, and replay-oriented logs.
Teaching and Leadership
Teaching Assistant, Python and AI Foundations
Summer 2025
Supported a course of more than 200 undergraduates building Avalon game bots, maintained the Tuvalon platform, and held help sessions for programming assignments and course projects.
Head of Academic Department, Student Union, School of Mathematical Sciences
Mar 2025 - Mar 2026
Led a 20-person team, organized academic talks and mock exams, and launched the department website pkusms.com.
