Beyond Execution: Static-Analysis Rewards and Hint-Conditioned Diffusion RL for Code Generation

In three linesEmpirical study of RL post-training for diffusion-based code generation. Authors propose execution-free rewards (static checking) and AST-hint-conditioned sampling to overcome the "capability cliff". Static checking improves DiffuCoder from 53.9 to 67.1 on HumanEval and reduces rollout time by 9.4%.

Summary generated by Claude — human-verified