🎓I am a first-year PhD student at Peking University. My advisor is Prof. Chi Zhang (张驰). My research focuses on generative evaluation and agentic reasoning in open-ended and dynamically generated environments.
My academic journey began at age 15 in the Wu Jianxiong Honor College at Southeast University, followed by a Master’s from UCAS, where I served as Student Union President. Prior to returning to academia, I spent four years as a Deep Learning Researcher at Baidu’s Autonomous Driving Department, leading the development of retrieval-based planning and self-supervised frameworks.
My recent work develops generative and augmented evaluation frameworks for frontier AI systems, including UniCode (ICML 2026) for code reasoning and MCU (ICML 2025 Spotlight) for open-ended game agents.
Research notes and essays.
Modern AI models that score perfectly on standardized benchmarks often fail in real-world applications. This post argues for generative evaluation: dynamic, diverse, and verifiable …
Published papers and preprints.
UniCode transforms seed programming problems into structurally augmented variants, builds robust test suites without human-written reference solutions, and diagnoses where code …
ICML 2025 Spotlight. A generative evaluation framework for open-ended game agents in Minecraft, featuring scalable task generation and multi-dimensional AutoEval.
MineStudio streamlines Minecraft AI agent development by integrating simulator, data, model, offline training, online training, inference, and benchmark modules into one …