R³L improves 3D layout generation with reliable spatial reasoning

Relative spatial relations are a compact way to describe spatial structure, but current Multimodal Large Language Models (MLLMs) often produce unreliable inferences when chained across multiple steps. The accumulation of errors from repeated reference-frame transformations leads to semantic and metric drift. To solve this, Zhifeng Gu, Yuqi Wang, and Bing Wang from the paper 'R³L: Reasoning 3D Layouts from Relative Spatial Relations' propose a general framework that breaks coupled relation chains via invariant spatial decomposition. This method ensures that spatial relationships remain consistent regardless of the reference frame used.

Additionally, R³L employs a consistent spatial imagination loop that iteratively imagines and revises layouts to promote self-consistency. Supportive spatial optimization uses global-to-local coordinate re-parameterization to ease pose optimization. Extensive experiments on various scene types and instructions show that R³L produces layouts that are both physically feasible and semantically aligned with user intent. The code is open-sourced on GitHub, and the paper has been accepted at ICML 2026.