Architect-Ant: Editable Automatic Furnishing of Architectural Floor Plans

Architect-Ant. Editable Automatic Furnishing of Architectural Floor Plans. By Fedor Rodionov and four others, from the King Abdullah University of Science and Technology and Miami University. Published on the thirtieth of September. No code repository is linked to it.

A furnished floor plan is useful to a lot of people: real-estate visualization, interior design, and architectural workflows all lean on one. Producing it automatically is harder than it sounds, for two reasons. Real-world data is limited. And a layout has to satisfy geometric and functional constraints that interact with each other, so a sofa that fits the wall can still block the door it was never meant to block. One way around this is iterative agentic inference, which the authors describe as costly. Their question is whether professional furnishing knowledge can instead be learned from real floor plans with a pretrained model, so that a constraint-aware layout comes out directly.

Two things are introduced together. The first is AntPlan, a curated dataset of five hundred and five real professional architectural floor plans. They carry dense furniture annotations spanning ninety-two object classes, across ten residential room categories. The second is Architect-Ant, the framework that generates the furniture layouts. Its central choice is the representation: layouts are written in an editable, coordinate-based DSL, a small language in which every piece of furniture is an object with a position.

Training happens in two stages. Supervised fine-tuning comes first, and that is where the model learns professional furnishing patterns from the plans. The second stage optimizes it with GRPO against a reward called the Layout Rule Score, or LRS. That score aggregates geometric and functional constraints derived from the professional plans themselves. It judges only the outcome, the finished layout, and prescribes no reasoning trace for how the model should get there. The supervision says what a good room looks like and leaves the route open.

Against diverse state-of-the-art baselines, Architect-Ant pairs low geometric violation rates with high functional completeness: the layouts break few geometric rules and still include what the rooms need. Qualitatively, they also more closely reflect real-world residential furnishing patterns, the kind of arrangement a professional would recognise.

Because the output is a layout in that coordinate language rather than a picture, it stays editable at the level of individual objects, and it can be converted into 3D scenes. Learned habits from real plans, then a rule score to keep them honest: a recipe aimed at skipping the expensive loop of an agent reasoning its way through every room.