LayoutDSL uses DSL actions to generate smarter interior layouts

Indoor scene layout generation has long been treated as a regression problem—models directly predict 3D bounding boxes and coordinates for furniture placement. According to the new paper, this oversimplification ignores structural elements like doors and windows and prevents models from learning the reasoning logic behind intelligent layout design. To fix this, Yuhao Lu and colleagues from five institutions propose LayoutDSL, a novel LLM-based framework that frames layout generation as a policy learning problem in a domain-specific language (DSL) action space. Each DSL action corresponds to an explicit, interpretable design decision (e.g., placing a bed relative to a window), rather than a raw geometric parameter.

The team also introduces 3D-FrontDSL, a large dataset pairing room-structure annotations with synthetic DSL action sequences for supervised fine-tuning. To go beyond imitation, they design reward functions based on interior design principles and physical plausibility, then optimize the policy using reinforcement learning. Experiments show LayoutDSL substantially improves spatial plausibility and design logicality over strong baselines and existing methods. The work aims to make AI-based interior design more reliable and interpretable, moving from black-box regression toward structured, verifiable reasoning. The paper is available on arXiv with the ID 2608.07547.