publications

This work studies how large language models perform on code-change-related tasks, including code review generation, commit message generation, and just-in-time comment update.

This work proposes ReCode, a reasoning-reinforced code generation framework that trains reasoning-process reward models and gates process rewards with strict execution outcomes.

This work introduces fault-guided fine-tuning for code generation by emphasizing error-sensitive code segments during training.

This work proposes self-explained keywords to help LLMs better understand low-frequency, problem-specific terms in code generation prompts.

This work studies whether LLMs can improve code generation without ground-truth supervision through label-free co-evolution of code and test generation.

This collaborative work studies local-cloud model cascading for low-latency and high-accuracy code completion.