SVoT: State-aware Visualization-of-Thought for Spatial Reasoning via Reinforcement Learning

In three linesSVoT is a reinforcement learning framework improving spatial reasoning in MLLMs by generating verifiable intermediate states and visualizations through transition reasoning chains. Trained with GRPO, SVoT achieves 65% absolute accuracy gain on extended domains (Pacman, Gather) requiring multi-object interactions.

Summary generated by Claude — human-verified