We adopt reinforcement learning to train LLMs to generate quality code with rewards derived from program analysis.

CoRe: a high-quality, multi-lingual benchmark for evaluating LLMs’ Code Reasoning capabilities with fundamental static analysis tasks.