DiffuCoder gains 24% on HumanEval via static-analysis rewards and hints

A new paper from Ouyang et al. systematically explores reinforcement learning post-training for diffusion language models in code generation, focusing on three axes: reward design, hint-conditioned sampling, and task difficulty. The core problem is that execution-based semantic rewards become too sparse on complex tasks, creating a 'capability cliff.' To address this, researchers evaluate execution-free alternatives like static checking (code linters, type checkers) and similarity-based rewards. They introduce hint-conditioned diffusion sampling that uses AST-based hints during training to guide exploration.

Experiments on HumanEval, MBPP, and LiveCodeBench show static checking as the strongest standalone execution-free reward, boosting DiffuCoder from 53.9 to 67.1 on HumanEval and from 14.9 to 15.5 on LiveCodeBench while reducing rollout time by 9.4%. They also find that moderate AST-based hinting is most useful on harder benchmarks, while similarity-based rewards work better on easier subsets. The work demonstrates that careful reward design and training-time hints can substantially improve diffusion-based code generation without needing to execute code.