UniFPDesign: Unified Floorplan and Scene Generation via LLMs

ACL ARR 2026 January Submission10336 Authors

Keywords: Floorplan Generation, Scene Synthesis, Large Language Models, Geometry-as-Code, Spatial Reasoning, Reinforcement Learning from Feedback, Instruction Tuning, Layout Design

Abstract: The fragmentation between topological floorplan partitioning and local scene synthesis has long hindered end-to-end automated residential design. We argue that a truly functional space requires a holistic reasoning chain that aligns immutable structural constraints with diverse user personas. To bridge this gap, we formalize the task of $\textbf{Constraint-Aware Unified Floorplan Design}$ and propose $\textbf{UniFPDesign}$, formulating the problem as ``Geometry-as-Code'' generation. To overcome the scarcity of aligned data, we introduce $\textbf{Persona-Guided Specification Inversion}$. By masking finalized layouts to reverse-engineer demands, this method utilizes structural invariants as implicit supervision and shifts from geometric rule-reversal to persona-centric synthesis. To eliminate spatial hallucinations, we develop a hierarchical CPT-SFT-RL strategy: CPT and SFT establish spatial syntax and instruction adherence, while GRPO ensures $\textbf{Geometric Habitability}$. Experiments on our proposed $\textbf{UniFPDesign-2K}$ benchmark demonstrate superior performance in integrating structural integrity with functional placement.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: Vision and Language, Language Grounding, Spatial Reasoning, Code Generation, Large Language Models, Reinforcement Learning

Contribution Types: NLP engineering experiment, Data resources

Languages Studied: English, Python (Geometry-as-Code)

Submission Number: 10336