{
  "id": 7586,
  "url": "https://arxiv.org/abs/2603.06459v1",
  "title": "Do Foundation Models Know Geometry? Probing Frozen Features for Continuous Physical Measurement",
  "summary": "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-training deficit rather than a representational one. Training objective determines accuracy more than architecture: five encoders",
  "authors": "Yakov Pyotr Shkolnikov",
  "category": "research",
  "topics": null,
  "orgs": null,
  "regions": null,
  "published_at": "2026-03-06T16:48:27.000Z",
  "fetched_at": "2026-07-14T16:33:21.047Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/7586",
  "original_url": "https://arxiv.org/abs/2603.06459v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}