Search papers, labs, and topics across Lattice.
This paper introduces P4IR, a two-stage framework that enhances large language model (LLM) performance in automated code compliance (ACC) by first applying supervised fine-tuning (SFT) to embed domain knowledge, followed by Group Relative Policy Optimization (GRPO) to refine the generated code representations. The framework demonstrated significant improvements, achieving reductions of up to 23.8% in tree edit distance and 38.6% in token-level Levenshtein distance compared to SFT baselines, while outperforming leading LLMs in both code structure and semantics in a zero-shot setting. Additionally, the GRPO stage contributed to a statistically significant decrease in false positives, indicating a more reliable ACC system.
P4IR reduces code compliance errors by up to 38.6% while outperforming top LLMs in accuracy and reliability.
Large language model (LLM)-based approaches for automated code compliance (ACC) of building regulations are prone to generating incorrect and hallucinated computer-processable rules. This paper introduces P4IR, a two-stage framework that uses supervised fine-tuning (SFT) to instill domain knowledge in an LLM, followed by Group Relative Policy Optimization (GRPO) to improve the accuracy of the generated intermediate representations in the form of high-level code skeletons. The framework achieved reductions of up to 23.8% and 38.6% in tree edit distance and token-level Levenshtein distance respectively, relative to the SFT baselines. Comparative analysis demonstrates that this approach in a zero-shot setting outperforms leading LLMs in both code structure and semantics, specifically Claude Opus and Sonnet 4.5, GPT-5.2, Qwen-3-Max, and GLM-4.7, evaluated via few-shot prompting. Additionally, the GRPO stage produced a small yet statistically significant reduction in false positives. By combining SFT with GRPO to optimize directly for domain-specific objectives, this approach offers a path toward more accurate and reliable LLM-based ACC systems.
NUS