Jia Li Linsay Li

My research connects code large language models with classical program analysis and reinforcement learning, spanning benchmarks, agentic systems, and training methods for real, repository-level, industrial-scale software. I was a research intern at WeChat, where several of these systems were deployed in production, and later at Tencent Hunyuan, working on post-training for long-horizon multi-turn agents. I am now with the Dots foundation model team at Xiaohongshu.

Research

I study how code agents use program structure, combine evidence across files, and learn from execution feedback in real software systems.

Vision: Self-Evolving Software Systems

My long-term goal is to build software agents that improve through use, learning from development, testing, and production feedback.

1. Plan / Code Change context & program structure
2. Build / Test Compilation errors & test feedback
3. Operate / Monitor Crash scenes & runtime logs
4. Feedback / Learn Repair outcomes & learnable reward

Next direction: turn sparse feedback from long-running software tasks into reliable training signals, and measure whether agents improve on new tasks.

Experience & Education

• 2026.09 – Present

  Talent Program Intern · Dots (Xiaohongshu) ACE Top Intern

  Foundation Model

  Large-scale data synthesis and agent post-training.

  dots3-note-prev

• 2026.5 – 2026.09

  Talent Program Intern · Hunyuan (Tencent) Qingyun Program

  Foundation Model Department

  Online data flywheel and reward-model preference alignment for multi-turn decision agents.

  offline PPOon-policy GRPO

• 2025 – 2026

  Research Intern · WeChat (Tencent)

  Technical Architecture Department

  Industrial crash diagnosis agents, and agentic RL for code repair deployed in production.

  weak feedbackmulti-turnlong-horizon

• 2023 – 2027

  Ph.D. in Computer Science · The Chinese University of Hong Kong

  Advised by Prof. Michael R. Lyu

• 2018 – 2022

  B.Eng. in Software Engineering · Sun Yat-sen University

Publications

First author on all listed works.

1. 

   Signal Reshaping for GRPO in Weak-Feedback Agentic Code Repair

   Preprint

   38.5% → 53.5% compile-and-semantic accuracy

   GRPO training improves code repair over the base model's zero-shot result by learning from compilation and semantic feedback.

   arXiv

2. 

   Holmes: Multimodal Agentic Diagnosis for Mixed-Language Mobile Crashes at Industrial Scale

   FSE 2026, Industry Track

   87.6% function-level localization accuracy
   ~40k production runs

   Diagnoses crashes in WeChat's mixed-language codebase by combining retrieval, code exploration, and runtime evidence.

   ACM

3. 

   From Laboratory to Real-World Applications: Benchmarking Agentic Code Reasoning at the Repository Level

   ACL 2026 Main Oral

   Reveals gaps in reasoning across files.

   RepoReason tests whether code agents can combine evidence across a repository to verify program behavior.

   arXiv

4. 

   VulKey: Automated Vulnerability Repair Guided by Domain-Specific Repair Patterns

   FSE 2026

   Improves repair accuracy on PrimeVul and Vul4J.

   Uses patterns from known vulnerabilities to guide LLM repairs across programming languages and models.

   arXiv

5. 

   MioHint: LLM-Assisted Request Mutation for Whitebox REST API Testing

   ICSE 2026

   Guides API testing beyond coverage plateaus.

   Uses program structure to guide LLM-generated API requests when automated testing stops making progress.

   arXiv

6. 

   ComBench: A Repository-level Real-world Benchmark for Compilation Error Repair

   Preprint

   Exposes the gap between compilable and correct repairs.

   Reproduces real C/C++ compilation failures from GitHub CI to evaluate whether code agents produce correct patches.

   arXiv