Spatial Harness grounds VLM spatial reasoning in physics-engine states
Original title: Spatial Harness: Grounding Spatial Reasoning in Executable World States with Physics Engines
On MindCube it raises Qwen3.5-27B from 52.70% to 71.14%, and to 98.25% when poses are annotated.
Researchers introduce Spatial Harness, which turns scenes into explicit executable world states so vision-language models can check spatial reasoning against a physics engine. A Perceiver places objects and cameras in a shared frame; a Reasoner then applies viewpoint changes using computed measurements and rendered views. Evaluations cover MindCube, SAT, MMSI-Bench, and SpatialRGPT-Bench with open- and closed-weight VLMs. Code is public.
Unconfirmed: Accuracy gains are author-reported and not independently verified here.
- MindCube score for Qwen3.5-27B rises from 52.70% to 71.14%.
- That result beats pySpatial by 7.52 percentage points.
- With identical annotated poses, accuracy is 98.25% versus 71.81% for internal reasoning.
- Code is at github.com/tmlr-group/Spatial-Harness.
- Products
- Spatial Harness
- AI models
- Qwen3.5-27B
From the original article
1 min read at huggingface.co
Spatial reasoning requires understanding how objects and observers are arranged in a scene and how their relations change with viewpoint or movement. Vision-language models (VLMs) typically perform both steps through model-internal reasoning, forming an implicit scene interpretation from partial views and imagining the consequences of spatial transformations.
By Chentao Cao, Zhanke Zhou, Sangni Duan · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.