Spatial Harness Grounds VLM Spatial Reasoning in Executable Physics-Engine World States
TMLR-Group-HF · hf · 2026-11-10
TMLR Group (HKBU) and ByteDance Seed introduce Spatial Harness, which grounds vision-language models' spatial reasoning in executable world states rather than implicit internal reasoning. A Perceiver fuses scene constraints from the question with specialized visual estimates to place objects and cameras in a shared coordinate frame, instantiating the state in a physics engine. A Reasoner then simulates viewpoint changes, using measurements and rendered views to guide answers—ensuring imagined transformations actually satisfy spatial constraints.
More from Research
- Hugging Face open-sources 14.8B DNA model Carbon-X, second only to GPT-6.1 Sol on GenomeQA — LoubnaBenAllal1 · 2026-11-10
- Agent swarm experiment: prompt name order and model choice decide who leads — bibryam · 2026-11-10
- New preprint frames Bayesian experimental design under misspecification as a max-min game — canaesseth · 2026-11-10
- François Fleuret asks: what are the key modern features of SOTA transformers? — francoisfleuret · 2026-11-10
- NBER paper: a third of significant development economics findings may be false — gleech · 2026-11-10
- Team fine-tunes Dream-1 with just 50 examples for a USB-unplugging robot task — IsaiahBallah · 2026-11-10