Reinforcement learning to improve large language model-based automated code compliance systems

The paper introduces P4IR, a two-stage framework designed to enhance the accuracy of large language model (LLM)-based automated code compliance systems. It employs supervised fine-tuning (SFT) to integrate domain knowledge, followed by Group Relative Policy Optimization (GRPO) to refine the generated code skeletons, achieving reductions of up to 23.8% in tree edit distance and 38.6% in token-level Levenshtein distance compared to SFT baselines. This approach demonstrates superior performance over leading LLMs like Claude Opus and GPT-5.2 in zero-shot settings, indicating its potential to improve the reliability of LLMs in generating accurate code representations for compliance tasks.