LLM-Based Floor Plan Generation with Reinforcement Learning

An innovative AI system enhances large language models (LLMs) by utilizing real floor plans and implementing reinforcement learning with verifiable rewards (RLVR) to ensure compliance with topological and numerical requirements. Current generative methods prioritize connectivity but overlook numerical factors such as room sizes and areas. The suggested approach creates text-based floor plans that fulfill both connectivity and numerical standards, surpassing current models. New metrics for constraint adherence are established to systematically evaluate conformity with user-specified constraints. This research fills a significant void in automated floor plan design intended for professional applications.