Does VGGT Know What It Does Not Know? Frozen-feature probe hits 93.5% AUC
kwangmoo_yi · x · 2026-10-28
arXiv paper 2610.35806 asks whether feed-forward 3D reconstruction model VGGT implicitly knows when its camera poses are wrong. On HardMatch — 1,000 deliberately difficult image pairs where VGGT fails on 35% — a linear probe on frozen features predicts pose failure at 93.5% ROC-AUC, beating VGGT's own point-map confidence (79.7%) and a classifier on all outputs (77.7%). The probe catches failures VGGT is confident about, enables downstream abstention, and transfers without retraining to rotation failures on PhotoTourism and ScanNet-1500. The signal arises partway through the layers fusing the two views and holds for π³ as well.
Related event: Study Finds VGGT Internally Knows When Its 3D Reconstructions Fail(3 posts)→
More from Embodied
- RAX lets a real robot explore for 2 hours and turn learned skills into reusable code — no demos needed — du_yilun · 2026-10-28
- OpenTac tactile sensors demoed at CoRL 2026 among 17 robot demos — _krishna_murthy · 2026-10-28
- Hybrid LLM agent scores 83.2%, beats nine VLA policies on new SO-101 robot arm benchmark — Truman Hickok · 2026-10-28
- cixliv: robot fighting may be the first realized market where bots beat average humans — cixliv · 2026-10-28
- Can the Hiwonder MentorPi T1 run fully offline as a robot pet? Pi 5 questioned — oulu2006 · 2026-10-28
- Testing Mint Modeler to Build the Menlo Research Robot — freelerobot · 2026-10-28