Software Engineering & Artificial Intelligence
I am Xin Zhou (周鑫), a Research Scientist at the School of Computing and Information Systems (SCIS), Singapore Management University (SMU), working with Professor David Lo. I received my Ph.D. in Computer Science from SMU in January 2025 under his supervision. Before that, I obtained my bachelor’s degree from Beijing Institute of Technology.
My research lies at the intersection of Software Engineering (SE) and Artificial Intelligence (AI), with the goal of making code LLMs and coding agents more capable, reliable, and useful for software development. Building on my earlier work on code representation learning and model generalization, my research now centers on three directions: (1) Agentic Software Engineering, developing techniques to improve coding agents’ core capabilities in requirement clarification, solution planning, program analysis, and issue resolution, while using trajectory analysis to identify bottlenecks and guide improvements; (2) Reliable Evaluation and Generalization, developing evaluation methods that reflect human judgments, account for benchmark contamination, and assess performance beyond training distributions, while extending evaluation from individual software development tasks to long-horizon, multi-step workflows; and (3) AI for Software Security and Reliability, advancing vulnerability detection, automated repair, and software quality assurance through program context and semantic reasoning.
This research has led to 30 publications, including 11 as first author, with work appearing in leading venues such as ICSE, FSE, ASE, TSE, TOSEM, and ACL.
I received the SMU Presidential Doctoral Fellowship in Computing (2024) and was recognized as a Distinguished Junior Program Committee Reviewer at MSR 2024. I also contribute to the research community through conference program committees, including the ICSE 2026 and ICSE 2027 Research Tracks, and reviewing for journals such as TSE, TIFS, and Empirical Software Engineering.
30 publications 1891 citations 20 h-index. Open to collaboration.
Ph.D. in Computer Science
B.S. in Automation Engineering
Improving coding agents’ core capabilities in requirement clarification, solution planning, program analysis, and issue resolution, using trajectory analysis to identify bottlenecks and guide improvements.
Coding agents · Core capabilities · Trajectory analysisEvaluating code LLMs and coding agents through human judgments, benchmark contamination analysis, and generalization beyond training distributions, extending from individual development tasks to long-horizon workflows.
LLM-as-a-Judge · Generalization · Long-horizon evaluationDeveloping AI techniques for vulnerability detection, automated repair, and software quality assurance, combining program context, semantic reasoning, and learning from code.
Vulnerability detection · Automated repair · Software reliabilityReleased SWE-RPG, a unified issue-resolution benchmark for requirement clarification, implementation planning, and code generation by coding agents. SWE-RPG paper
Released TraceProbe, a framework for diagnosing coding agents through trajectory structure beyond resolve rate. TraceProbe paper
Two papers on vulnerability repair and secure coding agents appear at ACL 2026. SeCuRepair · SecureVibeBench
Our study of out-of-distribution vulnerability detection appears at ICSE 2026. Paper
Our hierarchical framework bridging bug localization and issue fixing appears in IEEE TSE. Paper
SE-Jury, our human-aligned LLM-as-ensemble-judge metric, appears at ASE 2025. Paper
Evaluating coding agents across requirement clarification, implementation planning, and code generation, with validated references for diagnosing where trajectories go wrong.
Combining explicit repair reasoning, semantics-aware reinforcement learning, and a difficulty-based training curriculum to improve vulnerability repair across repositories.
Combining specialized LLM judges with dynamic team selection to evaluate generated code, patches, and summaries in closer agreement with human judgments.
Investigating training-data leakage across 83 software engineering benchmarks and removing leaked samples to support more reliable evaluation of code models.
Broadening the context available to code models with longer inputs, code structure, and expert knowledge for vulnerability repair.
Learning fine-grained representations of code changes from old and new code and edit actions, with self-supervised pre-training.
Beijing Institute of Technology
Mathematical Contest in Modeling (MCM)
Get in touch
For research inquiries, you can reach me by email.