Tsinghua University · Computer Science
Agents, large language models, and evaluation on real-world tasks.
I study the capability boundaries of agents on real-world tasks, and how reasoning, feedback, and self-improvement in large language models can help agents take on more complex, long-horizon work.
01About
AI is moving from answering questions to completing tasks. My research began with reasoning and recommendation in large language models, and now spans Coding Agents, engineering optimization, and algorithm design. I study how to measure agent capabilities in environments close to real work, and how planning, feedback, and self-improvement become transferable abilities.
Education
Computer Science · Undergraduate
2023 — Present · Beijing
Advisor · Prof. Hongning Wang ↗
Core courses
02Selected Publications
Selected work on agent evaluation, long-horizon tasks, and LLM reasoning.
01arXiv · 2026
A long-horizon Coding Agent benchmark for whole-repository technology-stack migration, combining migration audits, functional tests, and adversarial validation. Only 5.4% of 520 runs passed every check.
02arXiv · 2026
Introduces a cross-platform recommendation market led by User Agents, showing that broader product access does not automatically translate into effective exposure. With user feedback, positively promotional pitches fall from 73–78% to 36–41%.
03arXiv · 2026
R2Rank maps chain-of-thought reasoning over a user–candidate pair to a relevance score, then optimizes end to end with a Plackett–Luce distribution and NDCG-based reinforcement learning to reduce positional bias in LLM ranking.
04NeurIPS 2026 · Poster
A benchmark for self-evolving agents on real engineering optimization tasks, measuring whether agents can improve engineering artifacts within a fixed interaction budget through a propose–execute–evaluate–iterate loop and executable feedback.
05arXiv · 2026
A benchmark for LLM agents’ algorithm-design ability: agents explore and rewrite training algorithms in an open environment, then train and evaluate from scratch in an independent environment to separate algorithmic innovation from tuning.
06arXiv · 2026
Introduces a benchmark for coding agents to reconstruct scientific flow diagrams as native, editable PowerPoint slides. Across 500 tasks and 31 configurations, the best system reaches only 67.80, revealing persistent challenges in semantic and visual correctness.
07arXiv · 2026
A coding agent evolves reusable Skills and Lessons through feedback in a single Playground scene, then freezes them for diverse new scenes. On RoboDojo-Sim, EmbodiedEvo reaches an average score of 44.48 and a 37.82% success rate, surpassing all officially verified results.
03Beyond Research
Advanced curriculum reform and teaching feedback for the age of AI.
The only third-year Computer Science volunteer representative serving at the commemorative event.
Organized graduate-school sharing sessions and maintained student support services.
Participated in supervision and feedback on student affairs.
Organized joint theme activities across three schools; the class Youth League branch was recognized as a university-level A-grade branch.