Deep learning models can internally represent knowledge about their own failures even when they don't explicitly report uncertainty—this hidden knowledge can be extracted via probing and used to improve system reliability.
This paper investigates whether VGGT, a 3D reconstruction model that predicts camera poses without reporting uncertainty, actually knows when its predictions are wrong. Using a linear probe on frozen model features, researchers find the model can predict its own failures with 93.5% ROC-AUC on difficult image pairs, significantly better than the model's own confidence scores.