I work at the intersection of AI, software engineering, and systems. My research centers on agent evolution and recursive self-improvement (RSI): AI agents that keep getting better at real software work by learning from their own experience. A large part of this is LLM training, especially agentic reinforcement learning, on executable and verifiable environments at scale.
What sets my work apart is where it starts: the complexity of real products. Industrial codebases are huge, heterogeneous, and constantly evolving, with internal toolchains, build systems, and task types that public benchmarks never see. I model that complexity as scientific problems, then carry the results back into practice. I pioneered turning industrial product repositories and product tasks into executable data and environments for evaluating, tuning, and training coding agents, so that LLMs and coding agents attend to internal products while they are being trained and developed, and can be optimized for them.
Before Huawei, I spent eight years at Microsoft DKI (Data, Knowledge, Intelligence). There I led SWE-bench-Live and RepoLaunch (RepoLaunch builds agentic training environments for frontier open models such as GLM-5), contributed to UFO, TaskWeaver, and Agent Lightning, and worked with 10+ product teams to ship 20+ research results into products, including AIOps technologies in the fundamental cloud services behind Azure and Microsoft 365. I am also an adjunct master’s supervisor at the School of Computer Science, Fudan University. I received my PhD from The Chinese University of Hong Kong under Prof. Michael R. Lyu.
At the intersection of AI, software engineering, and systems: two core directions, built on one foundation, carried into three application scenarios.
Agents that do not stop improving at release. They learn from their own trajectories and failures, and they are trained against the products they will actually work on, not only against public benchmarks.
Closing the loop between acting and improving. Agents diagnose their own failures, revise their own harness, tools, and strategies, and generate the tasks and verification signals for the next round of improvement. The goal is improvement that compounds instead of stopping at a fixed model.
Training coding models and agents with reinforcement learning inside real agent harnesses, on executable, verifiable environments at scale. This covers environment construction, rewards and verification, trajectory data, and training frameworks. Agent Lightning v1.0 lifts Qwen3.5-9B on SWE-bench Verified from 41.8% to 56.4% with 6K examples.
Both directions need a steady supply of executable tasks with reliable verification. I turn product repositories, product tasks, and developer history into exactly that, so the same infrastructure serves evaluation, tuning, and training.
My research style: start from what makes a real product hard, state it as a research problem, and take the answer back into production.
Where the methods are applied and tested.
Issue resolution, test generation, code translation, dependency upgrades, and repository understanding, on both open-source and industrial code.
Desktop AgentOS and multi-device agent systems, plus a widely cited survey of LLM-brained GUI agents.
Incident detection, triage, root-cause analysis, logging, and monitoring for hyperscale cloud, deployed at Microsoft.
Environments, benchmarks, frameworks, and systems used outside the lab: in frontier-model training and evaluation, by large open-source communities, and in hyperscale production.
“We employ an environment setup pipeline based on the RepoLaunch framework that scales the construction of executable environments from real-world SWE issues.”
— GLM-5 Technical Report, Zhipu AI / Z.ai (arXiv 2602.15763)
RepoLaunch is the first agent that resolves dependencies, compiles code, and extracts test results for any language on any OS: C/C++, C#, Python, Java, JS/TS, Go, and Rust, on Linux and Windows. It was used to build 10,000+ Docker environments for large-scale agentic training.
Qwen3-Coder’s release reports SWE-bench-Live next to Kimi-K2, DeepSeek-V3, Claude Sonnet 4, and GPT-4.1.
“Our original Agent Lightning introduced a disaggregated architecture that connects arbitrary agents to RL training through an LLM endpoint proxy, an approach later adopted by frameworks such as verl Uni-Agent, AReaL 2.0, slime, and Polar.”
— Agent Lightning v1.0 (arXiv 2608.17528)
20+ results shipped with 10+ Microsoft product teams (Azure, Microsoft 365, Windows, DevDiv, Copilot): AIOps across the incident lifecycle, and resource-management algorithms for Azure and Microsoft 365 that save hundreds of millions of dollars a year. 10+ patents.
A short selection, chosen for adoption, citations, or top-venue publication. Led marks projects I led.
A continuously updated issue-resolution benchmark, built by an automated pipeline. RepoLaunch is an agent that sets up, builds, and tests any repository, in any language, on Linux or Windows, turning real issues into verified, reproducible Docker tasks with no human in the loop. The benchmark now has MultiLang (1,077 tasks · 8 languages) and Windows splits, and the pipeline reaches ≥98% build/test success at 82% lower LLM cost.
Product repositories become a renewable source of agent tasks. Change2Task converts merged changes into verified tasks on healthy modern revisions (bug fix, feature addition, test generation, API migration, security repair), reaching 79.6% verified construction and 29.2% more tasks than a PR-based baseline. A companion line of work generates continuous benchmarks for enterprise-scale agents.
A lightweight framework for harnessed agentic RL: the deploy-time agent harness owns the interaction loop and takes part in post-training directly. It addresses retokenization, sample merging, advantage calculation, and loss normalization, and ships a full coding-agent RL pipeline: 41.8% → 56.4% on SWE-bench Verified for Qwen3.5-9B. At Microsoft I drove its large-scale distributed deployment on internal clusters, with rollout and training decoupled.
Automatic debugging for LLM multi-agent systems, a step toward agents that repair themselves. DoVer segments the trace, forms failure hypotheses, intervenes at the orchestrator, then replays to validate the fix.
UFO was a pioneering UI-focused agent for Windows. UFO² turned it into a multi-agent Desktop AgentOS (HostAgent + AppAgents, hybrid GUI–API actions), and UFO³ connects agents across devices. Related work includes the survey of LLM-brained GUI agents (230+ citations) and a study of how API and GUI agents diverge and converge (ICML 2025).
One of the first LLM systems for root-cause analysis of cloud incidents: incident-specific handlers collect diagnostics, and an LLM predicts and explains the root cause. Evaluated on a year of real Microsoft incidents.
Papers chosen for top-venue publication, citations, or adoption. See the full list or Google Scholar.
Software Engineering Application Technology Lab, Huawei
Agent evolution and RSI, LLM training and agentic RL, and product-grounded environments for coding agents.
Microsoft DKI (Data, Knowledge, Intelligence; formerly part of Microsoft Research Asia)
Led SWE-bench-Live and RepoLaunch. Contributed to Agent Lightning, UFO, TaskWeaver, and a line of AIOps and resource-management systems shipped with 10+ product teams (Azure, Microsoft 365, Windows, DevDiv, Copilot).
School of Computer Science, Fudan University
School of Computer Science, Fudan University
Co-advisor: Prof. Xin Wang
The Chinese University of Hong Kong
Software Analytics Group, Microsoft Research Asia
Star of Tomorrow Award.
Computer Science, Duke University
Host: Prof. Kishor S. Trivedi
The Chinese University of Hong Kong
Advisor: Prof. Michael R. Lyu
Mobile app performance diagnosis (DiagDroid, FSE’16), cloud service deployment, and software security.
Fudan University
Top 5%; first-class scholarship three years in a row.
Our research team is growing. If you want to work on agent evolution, agentic RL, and coding agents that meet the full complexity of real products, I would like to hear from you.