Spatial-Interactor: Learning Spatial Reasoning through Interaction with the Observable Physical World
The paper trains vision-language models to track spatial state changes across interaction trajectories, not just answer static spatial questions.
Spatial-Interactor uses a three-level curriculum: passive world-state transitions, active self-state transitions, and long-horizon interaction trajectories. The authors build LSI-108K from simulated and real trajectories to supervise those stages. Training pairs SFT for local transition modeling with on-policy distillation, where a teacher branch with segment-level transition descriptions guides the student’s chain-of-thought over longer paths. The paper reports consistent gains across multiple VLMs and spatial benchmarks. HF Daily Papers' note
Spatial-Interactor uses a three-level curriculum: passive world-state transitions, active self-state transitions, and long-horizon interaction trajectories. The authors build LSI-108K from simulated and real trajectories to supervise those stages. Training pairs SFT for local transition modeling with on-policy distillation, where a teacher branch with segment-level transition descriptions guides the student’s chain-of-thought over longer paths. The paper reports consistent gains across multiple VLMs and spatial benchmarks. HF Daily Papers' note
score 5