Spatial Harness: Grounding Spatial Reasoning in Executable World States with Physics Engines
With identical annotated object and camera poses, Spatial Harness reached 98.25% accuracy versus 71.81% for model-internal reasoning.
The system places objects and cameras in a shared coordinate frame, then uses a physics engine to execute viewpoint changes. Computed measurements and rendered views guide the model’s answers. On MindCube, it raised Qwen3.5-27B’s accuracy from 52.70% to 71.14%, beating pySpatial by 7.52 percentage points. The authors also evaluated it on SAT, MMSI-Bench and SpatialRGPT-Bench using open- and closed-weight vision-language models.
HF Daily Papers' note
The system places objects and cameras in a shared coordinate frame, then uses a physics engine to execute viewpoint changes. Computed measurements and rendered views guide the model’s answers. On MindCube, it raised Qwen3.5-27B’s accuracy from 52.70% to 71.14%, beating pySpatial by 7.52 percentage points. The authors also evaluated it on SAT, MMSI-Bench and SpatialRGPT-Bench using open- and closed-weight vision-language models.
HF Daily Papers' note
score 5