An evaluation guided multiagent framework for generating floor plans that comply with municipal building regulations

Abstract Generative models often produce perceptually plausible outputs that violate hard domain rules. In architectural layout generation, balancing geometric intuition with regulatory compliance remains a persistent challenge, as current frameworks routinely fail to conform to local building laws. This paper proposes LayoutGen, a multi-agent neuro-symbolic framework coupling probabilistic generation with deterministic verification in an evaluation-guided refinement loop. LayoutGen employs a four-phase retrieval-driven pipeline (spatial reasoning, coordinate mapping, vector rendering, and iterative refinement) powered by a dual retrieval engine leveraging architectural precedents and legislative laws. An Evaluation Engine scores the generated 2D Scalable Vector Graphics (SVG) layouts and returns a graded, itemised report. Every scored quantity is computed by deterministic geometric checks over the generated coordinates against a structured rule repository; the Large Language Model (LLM) is confined to spatial planning, coordinate proposal, and the natural language feedback that drives refinement, together with an advisory Vaastu commentary that does not enter the grade. Evaluated on single-storey residential configurations drawn from the RPLAN and MSD datasets, spanning studio to 3BHK typologies and represented as axisaligned rectangular rooms, the full pipeline reduces detected violations from 3.7 to 0.3 per plan across n =200 runs, a 92% reduction relative to an unretrieved, unrefined single-shot baseline using the same backbone model. Violations are counted across the three error categories the Evaluation Engine checks: topological, geometric, and the encoded dimensional provisions of minimum room area, minimum room width and critical width. The regulatory component of that count covers only provisions decidable from single-floor room geometry; setbacks, floor area ratio, egress, staircase geometry, parking and ventilation are present in the repositories but checked outside the generation loop. Component ablation isolates the contribution of each pipeline stage, and the same architecture runs unchanged across three Indian municipal rule repositories (BBMP, CMDA, GHMC) through structured rule replacement alone, without retraining or prompt redesign. These three regimes express their dimensional provisions in a common form, so this establishes portability within that family rather than generality across regulatory systems. The framework also integrates Server-Sent Events (SSE) for real-time design streaming.