Latent representation probes let VLMs abstain when OCR is uncertain

The researchers propose Latent Representation Probing (LRP), which trains lightweight probes on VLM hidden states or attention patterns to make them abstain when OCR is uncertain. Three probe designs improve abstention accuracy by 7.6% over the best baseline on four benchmarks covering image and video modalities. The probes generalize across uncertainty sources and datasets, and the best signal comes from intermediate layers rather than the last layer.