ISSTA 2026Program Repair
To Run or Not to Run
Analyzes the cost and benefit of code execution in LLM-based repair, arguing that execution feedback is a costly resource rather than a default action.
About
I am a Ph.D. student at Beihang University, working in the SMAT Laboratory under Prof. Li Li, co-supervised by Mingyi Zhou.
My research focuses on AI for Software Engineering (AI4SE), combining large language models, program analysis, and verifiable software-engineering environments to build reliable coding agents and ecosystem toolchains. The work spans execution feedback, repository-scale benchmarks, change impact analysis, OpenHarmony quality assurance, and low-resource programming languages — all around one core question: how to make AI genuinely reliable for software engineering practice.
Motivation
LLMs already generate code well, but real software engineering is not only about writing a patch: an agent must understand repositories, select evidence, control execution cost, maintain change history, and survive verification in a specific ecosystem. My central question is how to combine LLM generation, the determinism of program analysis, and real software-engineering environments into reliable coding agents and ecosystem toolchains.
In coding agents, this means analyzing when execution feedback is worth its cost (To Run or Not to Run) and using lightweight static structure to stabilize repository navigation (CodeAnchor). In the OpenHarmony ecosystem, it means combining domain rules, static analysis, LLMs, and real feedback for repair, migration, and performance analysis (HapRepair, Phantom Rendering Detection). On the evaluation side, it means systematically examining what code knowledge LLMs actually understand and how AI4SE tools should be measured.
Education
- 2024 – present
- Beihang University · Computer Science · M.S.–Ph.D. program
- 2019 – 2024
- Beihang University · Software Engineering · B.Eng.
News
- Selected for the 2026 PhD Student Young Talent Development Program(Chinese)— Moving forward with renewed encouragement to pursue meaningful research at the intersection of AI and software engineering.
- OpenHarmony Bench is on arXiv: successful compilation does not guarantee task completion(Chinese)— An app-level coding-agent benchmark I contributed to during my internship, covering new features, specification-driven implementation, and real bug fixes in OpenHarmony.
- AtomicCommitBench is on arXiv: coding agents should deliver more than a single diff(Chinese)— Reorganizing squashed patches into atomic commit histories that can be replayed, reviewed, and selectively reverted.
- Introducing SkillLens: an open-source framework for visualizing and optimizing skills for Codex and Claude Code— Aligning SKILL.md constraints with agent traces to reveal followed, violated, and ignored instructions and suggest improvements without bloating the skill.
- RepoRescue is on arXiv: rescuing old repositories from ecosystem drift(Chinese)— Studying whether LLM agents can repair source code to restore repositories affected by changes in their software ecosystems.
Selected Work
ISSTA 2026Code Agents
CodeAnchor
Studies how lightweight static structure acts as deterministic anchors, improving localization stability, reproducibility, and navigation efficiency in real repositories.
FSE 2026Performance
Phantom Rendering Detection
Identifies unnecessary UI computations so developers can find hidden performance bottlenecks and improve mobile app smoothness.
ICSE 2026LLM Security
MazeBreaker
Proposes a multi-agent reinforcement learning framework that dynamically evaluates jailbreak vulnerabilities in LLM safety defenses.
Timeline
- Late 2023Started research as an undergraduate and joined the SMAT Laboratory.
- 2023.11First submission — Open-Source AI-based SE Tools, a survey of the AI4SE open-source ecosystem.
- 2024First paper accepted at TOSEM — a panoramic survey covering 100+ open-source tools.
- 2024.09Graduated and began the combined M.S.–Ph.D. program; started collaborations with Huawei, BTH, and SMU.
- 2025.03HapRepair accepted at FSE Industry 2025 — automated repair for the OpenHarmony ecosystem, from paper to engineering.
- 2025.06MazeBreaker accepted at ICSE 2026 — adaptive security evaluation with multi-agent reinforcement learning.
- 2025 – 2026Research expands: Phantom Rendering (FSE 2026), Cangjie fine-tuning (EMSE 2026), and several submissions on repository-scale compatibility maintenance, change impact analysis, and OpenHarmony migration.
Research Directions
Four threads sharing one question: how to make coding agents both reliable and verifiable in real software-engineering environments. Click to expand.
Research Themes
Coding AgentsProgram RepairExecution FeedbackRepo-Scale BenchmarksChange Impact AnalysisAtomic CommitsOpenHarmonyStatic AnalysisCode UnderstandingLow-resource Languages
Accepted Papers
All papers →- To Run or Not to Run: Analyzing the Cost-Effectiveness of Code Execution in LLM-Based Program RepairISSTA 2026PDF
- How Much Static Structure Do Code Agents Need? A Study of Deterministic AnchoringISSTA 2026PDF
- Phantom Rendering Detection: Identifying and Analyzing Unnecessary UI ComputationsFSE 2026PDF
- MazeBreaker: Multi-Agent Reinforcement Learning for Dynamic Jailbreaking of LLM Security DefensesICSE 2026PDF
- HapRepair: Learn to Repair OpenHarmony AppsFSE Industry 2025PDF
- Effective Fine-tuning for Low-resource Languages: A Case Study of CangjieEMSE 2026PDF
- Open-Source AI-based SE Tools: Opportunities and Challenges of Collaborative Software LearningTOSEM 2024PDF
- Exploring Code Analysis: Zero-Shot Insights on Syntax and Semantics with LLMsTOSEM 2026PDF
Open Source
HapRay
Performance analysis tool for OpenHarmony apps, detecting Phantom Rendering issues in UI.
FSE 2026 · GitCode
HomeCheck
Static checker and auto-repair tool for OpenHarmony apps; core module of HapRepair.
FSE Industry 2025 · GitCode
HomeTrans
Android → OpenHarmony migration tool, helping apps enter the HarmonyOS ecosystem.
Migration tool · GitCode
ResearchClaw
AI-powered research desktop app: literature management, smart reading notes, and idea generation.
Electron + TypeScript · GitHub
SkillLens
SKILL coverage and optimization framework for Codex / Claude Code agent traces.
Agent Skills · GitHub
Collaborations
Huawei
OpenHarmony / mobile app quality
Chi Chen
- HapRepair: OpenHarmony app repair
- Cangjie low-resource language fine-tuning
Han Hu
- Phantom Rendering detection for mobile UI performance
Bo Sun
- Phantom Rendering detection for mobile UI performance
Gang Fan
- Phantom Rendering detection for mobile UI performance
Topics: OpenHarmony app repair, Phantom Rendering / UI performance analysis, Cangjie low-resource language fine-tuning
Blekinge Institute of Technology (BTH)
LLM code understanding / AI4SE evaluation
Heartfelt thanks to Wei Ma for guiding me into research. When I first began, he patiently helped me with topic selection, paper reading, experiment design, and writing; much of my early research training was shaped by his guidance and support.
- MazeBreaker: multi-agent RL for LLM jailbreak evaluation
- HapRepair: LLM-guided repair for OpenHarmony apps
- Exploring Code Analysis: syntax and semantic probing with LLMs
- Open-source AI-based SE tools survey
Topics: LLM code understanding and semantic evaluation, MazeBreaker adaptive evaluation, AI4SE open-source ecosystem studies
Singapore Management University (SMU)
LLM agents / execution efficiency
Heartfelt thanks also to Zhensu Sun. After I became able to develop research ideas more independently, he continued to offer constructive advice on problem framing, experiment design, paper narrative, and limitation analysis, helping me learn how to refine early ideas into more complete and solid research work.
- EAGER: executing code as LLMs generate it
- To Run or Not to Run: execution cost-effectiveness in LLM repair agents
Topics: LLM coding agents, Execution cost and latency optimization, Code generation and repair evaluation