Tsinghua University
Sep. 2025 – Present
M.Eng. in Data Science and Information Technology
Advisors: Xinlei Chen and Chao Yu.
About Me
I am a 2025 master's student in Data Science and Information Technology at the Tsinghua Shenzhen International Graduate School, co-advised by Xinlei Chen and Chao Yu. I received my bachelor's degree from Beijing University of Posts and Telecommunications.
My research focuses on LLM agents and reinforcement learning. My current interests include recursive self-improvement (RSI), credit assignment, and world models; my earlier work covered multi-agent reinforcement learning and multi-UAV systems.
Publications
Full list on Google ScholarActionOPD: Distilling Decisions, Not Tokens
Online policy distillation for LLM agents at the semantic action level: aligns student decisions with black-box teacher-sampled actions without teacher token probabilities or a shared tokenizer, yielding 3.3–6.9 point gains on out-of-distribution benchmarks.
Verifiable Process Rewards for Agentic Reasoning
Turns verifiable task-solving oracles into action-level process rewards, providing dense feedback for long-horizon agentic tasks and outperforming outcome-only rewards (GRPO, VinePPO) in Sokoban, Sudoku, Minesweeper, and τ²-bench.
MARSHAL: Incentivizing Multi-Agent Reasoning via Self-Play with Strategic LLMs
An end-to-end RL framework where strategic LLMs learn multi-agent reasoning directly from game outcomes through cooperative and competitive self-play, with gains transferring to downstream multi-agent systems like MAD and AutoGen.
VS-Bench: Evaluating VLMs for Strategic Abilities in Multi-Agent Environments
A multimodal benchmark of 10 vision-grounded multi-agent environments (cooperative, competitive, and mixed-motive) evaluating 15 VLMs on perception, reasoning, and decision-making.
Scalable UAV Multi-Hop Networking via Multi-Agent Reinforcement Learning with Large Language Models
Combines multi-agent reinforcement learning with LLMs for decentralized coordination in scalable UAV multi-hop networks.
VolleyBots: A Testbed for Multi-Drone Volleyball Game Combining Motion Control and Strategic Play
A robot sports testbed where multiple drones cooperate and compete in volleyball under physical contact, integrating high-level strategic play with low-level agile flight control, with zero-shot sim-to-real deployment.
Work Experience
Education
Beijing University of Posts and Telecommunications
Sep. 2021 – Jun. 2025
B.Eng. in Information Engineering
Honors & Awards
- 2025Merit Student, Beijing University of Posts and Telecommunications (2024–2025)
- 2024Merit Student, Beijing University of Posts and Telecommunications (2023–2024)
- 2023National Scholarship (2022–2023)
- 2023Outstanding Student Leader, Beijing University of Posts and Telecommunications (2022–2023)
- 2023Outstanding Individual in Student Organizations, Beijing University of Posts and Telecommunications (2022–2023)