Linear probe predicts VGGT pose failures at 93.5% AUC, paper finds
kwangmoo_yi · x · 2026-10-28
The paper "Does VGGT Know What It Does Not Know?" shows that feed-forward 3D reconstruction model VGGT encodes knowledge of its own failures. On HardMatch (1,000 difficult image pairs where VGGT fails on 35%), a logistic regression probe on frozen features predicts pose failure at 93.5% ROC-AUC, versus 79.7% for VGGT's own point-map confidence. The signal emerges in layers fusing the two views, transfers without retraining to PhotoTourism and ScanNet-1500, and also holds for π³.
Related event: Study Finds VGGT Internally Knows When Its 3D Reconstructions Fail(3 posts)→
More from Research
- Turing Post weekly must-read papers: WorldMirror, VerifierBench, agent RL harness scaling — TheTuringPost · 2026-10-28
- Stanford AIMI releases CheXthought: 103K radiologist reasoning traces on 50K chest X-rays — mattlungrenMD · 2026-10-28
- New research agenda on 'inter-AI morality': how AI systems perceive and treat each other — dioscuri · 2026-10-28
- All Seven Millennium Prize Problems See Breakthrough Claims, Several Driven by AI Labs — haider1 · 2026-10-28
- Fine-tuning VLAs on TAMP data hits distribution mismatch; DATAFARM gets JAX port — tomssilver · 2026-10-28
- CMU L3 Lab kicks off CUA Week with multiple realtime agent research releases — wellecks · 2026-10-28