Vision-Language Models (VLMs) excel at interpreting 2D images but often struggle with spatial reasoning about the physical world. Researchers argue this stems from a dimensional mismatch: VLMs are trained on 2D projections, yet true spatial reasoning requires recovering latent 3D geometry and temporal continuity. To address this, they propose a 'divide and conquer' approach called FactoSR.
FactoSR is a factorized reinforcement learning framework that explicitly decomposes the monolithic problem of world-consistent reasoning into three orthogonal geometric sub-objectives: planar correspondence (XY), depth consistency (Z), and temporal reversibility (T). By optimizing these verifiable constraints within a unified policy learning mechanism, the framework transforms an ill-posed projection recovery problem into a series of tangible reasoning steps.
Extensive evaluations on multi-view and video benchmarks demonstrate substantial gains. FactoSR achieves a 5.9% improvement on VSI-Bench and a 4.5% improvement on All-Angles-Bench, showcasing the effectiveness of reinforcing explicit, factorized 4D consistency.
The paper is available on arXiv (2609.03729), and code is open-sourced on GitHub. The authors believe this approach is a critical step toward evolving VLMs into robust, world-aware reasoners.