Training and Reinforcement Learning for Code Generation
Fault-guided fine-tuning and reasoning-process reward learning for code LLMs.
This project line focuses on making code LLM training more sensitive to the real reasons why generated programs fail. FGit emphasizes error-sensitive segments during fine-tuning, while ReCode studies reasoning-process rewards for code generation and gates process feedback with strict execution outcomes.
The broader goal is to move beyond outcome-only supervision and design training signals that better align model behavior with correct, robust code generation. Recent follow-up work, ZeroCoder, further explores label-free co-evolution of code and test generation without ground-truth supervision.