AI RESEARCH
Do Foundation Models Know Geometry? Probing Frozen Features for Continuous Physical Measurement
arXiv CS.AI
•
ArXi:2603.06459v1 Announce Type: cross Vision-language models encode continuous geometry that their text pathway fails to express: a 6,000-parameter linear probe extracts hand joint angles at 6.1 degrees MAE from frozen features, while the best text output achieves only 20.0 degrees -- a 3.3x bottleneck. LoRA fine-tuning (r=16, 2,000 images) narrows this gap to 6.5 degrees, providing evidence for a pathway-