This is Ye, welcome to my page. I am a Lecturer (aka. Assistant Professor) in the Department of Computer Science at University College London. My research builds the next generation of coding agents that work like software engineers inside a real command-line environment, reasoning across an entire codebase and its tools to tackle tasks such as issue resolution, agent failure repair, and more. My group focuses on CLI coding agents: we design the agent harnesses they operate in, train them in realistic terminal-world environments, and study why they fail and how they can repair themselves. What I care about most is making these agents secure and cost-efficient, so that they are actually affordable in practice. We also maintain Awesome Code Agents, a curated list of cutting-edge coding-agent projects and research.

Prior to joining UCL, I worked as a Postdoctoral Researcher at Carnegie Mellon University with Prof. Claire Le Goues. I obtained my PhD from KTH Royal Institute of Technology, where I was fortunate to be supervised by Prof. Martin Monperrus and Prof. Benoit Baudry. I received my bachelor’s degree from Sichuan University.

Let’s connect: I’m always happy to talk research, swap ideas, or just say hi — whether we share interests or come from very different fields. Reach me at [email protected].

News

Projects

TerminalWorld: Benchmarking Agents on Real-World Terminal Tasks.

Zhaoyang Chu, Jiarui Hu*, Xingyu Jiang*, Pengyu Zou*, Han Li, Chao Peng, Peter O’Hearn, Earl T. Barr, Mark Harman, Federica Sarro, He Ye†.

Preprint.

TerminalWorld is a benchmark that evaluates AI agents on the real-world terminal workflows developers run every day; its novelty is an automated pipeline that mines real terminal recordings into reproducible, test-verified task environments that stay live as engineering practices evolve.

arXiv2605.22535 GitHub36 HuggingFace10438 Websiteterminalworld.ai

🎙️ TerminalWorld was featured on Last Week in AI (ep. #246).

ContextBench: A Benchmark for Context Retrieval in Coding Agents.

Han Li, Letian Zhu, Bohan Zhang, Rili Feng, Jiaming Wang, Yue Pan, Earl T Barr, Federica Sarro, Zhaoyang Chu, He Ye.

Preprint.

ContextBench is a benchmark that evaluates how coding agents perform multi-file context retrieval across a repository; its novelty is measuring the dynamics of retrieval — not just whether the right files are found, but the accuracy (context F1), efficiency, and cost of how agents gather that context.

arXiv2602.05892 GitHub149 HuggingFace1028 Websitecontextbench.github.io

Prometheus: Towards Long-Horizon Codebase Navigation for Repository-Level Problem Solving.

Yue Pan, Zimin Chen, Siyu Lu, Zhaoyang Chu, Xiang Li, Han Li, Yang Feng, Claire Le Goues, Federica Sarro, Martin Monperrus, He Ye.

Preprint.

Prometheus is a coding agent for repository-level problem solving that navigates large codebases over long horizons; its novelty is unifying embedding-based retrieval, structure-aware knowledge-graph navigation, and a working memory over the agent's trajectory, so it gathers the right context across many files and reasoning steps — reaching top results on SWE-bench Verified and SWE-PolyBench.

arXiv2507.19942 GitHub1000+