About Me

I am an undergraduate at Yuanpei College, Peking University, majoring in Data Science and Big Data Technology. I work with Prof. Bin Cui and Prof. Xupeng Miao at PKU, and Prof. Fan Lai at the University of Illinois Urbana-Champaign. My research focuses on efficient LLM algorithms and systems, especially reinforcement learning, post-training, and diffusion language models. I am also interested in recursive self-improvement and context management for long-horizon agents. I study how algorithm and system design can work together to improve training and inference efficiency without sacrificing model capability.

Education

Peking University, Yuanpei College

Sep 2023 – Present

B.S. in Data Science and Big Data Technology

GPA: 3.8/4.0   Major rank: 1/13   Overall rank: 43/355

Selected coursework: Machine Learning (99); Programming in Artificial Intelligence (99); Principles, Challenges, and Industrial Applications of LLMs (99); Practice of Programming in C & C++ (98); Data Structure and Algorithm (A) (95); Probability Theory and Statistics in Information Science (95); Directed Research in AI System (95); Advanced Algebra (93).

Teaching assistant: Probability Theory and Statistics in Information Science; Outdoor Exploration.

Tsinghua University High School

Sep 2017 – Jun 2023

High School Certificate

Research Experience

Efficient Diffusion Language Models

Feb 2026 – Present

University of Illinois Urbana-Champaign · Advisor: Prof. Fan Lai

Parallel generation, scalable training, and efficient serving. In BRISK-DLM, we study risk-aware proposal learning and prefix-conditioned selection for faster verified decoding.

Efficient Reinforcement Learning for LLMs

May 2025 – Present

Peking University · Advisors: Prof. Bin Cui and Prof. Xupeng Miao

Investigating rollout stragglers and response length explosion through DARTS and QLPO, combining system optimization with length-aware policy learning.

Context Management for Long-Horizon Agents

Feb – May 2026

Peking University · Advisors: Prof. Bin Cui and Prof. Xupeng Miao

Designed diagnostic tasks and evaluation protocols for MemClaw-Bench, studying memory capabilities in long-horizon tool-using agents.

Publications & Preprints

* Equal contribution

MemClaw-Bench: multi-episode tasks, context management, agent execution, and oracle-based diagnosis across eight memory capabilities

MemClaw-Bench: A Capability-Centered Benchmark for Context Management in Long-Horizon Agents

Siwei Chen, Xupeng Miao, Bin Cui

arXiv preprint, 2026

Evaluating constraint tracking, version updating, procedure transfer, source conflict resolution, task resumption, and staleness judgment under compressed context.

Projects

Usagi: ICS Course Helper

A retrieval-augmented course assistant with source tracing, predictive question generation, and learner state diagnosis. Used by 100+ students.

Programming & tools: Python, C++, Rust, PyTorch, CUDA.

Honors & Awards

  • National Scholarship (China’s highest-level undergraduate scholarship)2025–2026
  • Academic Innovation Award, Peking University2025–2026
  • Outstanding Student Model, Peking University2025–2026
  • SenseTime Scholarship (selected among 30 recipients nationwide)2026
  • Outstanding Research Award, Peking University2024–2025
  • Third-Class Scholarship, Peking University2024–2025
  • Merit Student, Peking University2023–2024
  • Second-Class Scholarship, Peking University2023–2024
  • Freshman Scholarship, Peking University2023–2024
  • Beijing Merit Student2023

Miscellaneous

I enjoy basketball, running, Ultimate Frisbee, and mountain climbing. I am a member of Shanying Club, Peking University's mountaineering association, and won second prize at the PKU Ultimate Frisbee Cup in 2024 and 2025.

Snapshots

Loading