FactoSR Factorizes 4D Properties for Improved Spatial Reasoning
September 2, 2026
FactoSR is a factorized reinforcement learning framework designed to overcome the 2D projection bottleneck in Vision-Language Models. It decomposes world-consistent reasoning into three orthogonal geometric sub-objectives to recover latent 3D geometry and temporal continuity.
HOW THIS AFFECTS YOU
●
researcherThis provides a framework to move VLMs beyond flat 2D interpretations toward true spatial reasoning.