factorized-learning

This paper introduces FactoSR, a factorized reinforcement learning framework that enhances spatial reasoning in Vision-Language Models by decomposing 4D properties into orthogonal sub-objectives, achieving significant performance boosts on multi-view and video benchmarks.