Reading Between the Lines: Abstaining from VLM-Generated OCR Errors via Latent Representation Probes

TL;DR

This work tackles the reliability gap of Vision-Language Models in OCR-rich Scene Text VQA by enabling abstention through internal model signals. The authors propose Latent Representation Probing (LRP), which trains lightweight probes on hidden states and attention patterns to detect uncertainty, with three probe designs and two latent-representation modalities. Across four STVQA benchmarks, LRP—especially ensemble probes on hidden states—consistently outperforms baselines, achieving up to a 7.6% improvement in abstention accuracy and demonstrating strong generalization to unseen uncertainty sources and datasets. A key finding is that optimal confidence signals emerge from intermediate layers, enabling robust, deployment-ready abstention without expensive fine-tuning or reliance on output probabilities, including for long video sequences.

Abstract

As VLMs are deployed in safety-critical applications, their ability to abstain from answering when uncertain becomes crucial for reliability, especially in Scene Text Visual Question Answering (STVQA) tasks. For example, OCR errors like misreading "50 mph" as "60 mph" could cause severe traffic accidents. This leads us to ask: Can VLMs know when they can't see? Existing abstention methods suggest pessimistic answers: they either rely on miscalibrated output probabilities or require semantic agreement unsuitable for OCR tasks. However, this failure may indicate we are looking in the wrong place: uncertainty signals could be hidden in VLMs' internal representations. Building on this insight, we propose Latent Representation Probing (LRP): training lightweight probes on hidden states or attention patterns. We explore three probe designs: concatenating representations across all layers, aggregating attention over visual tokens, and ensembling single layer probes by majority vote. Experiments on four benchmarks across image and video modalities show LRP improves abstention accuracy by 7.6\% over best baselines. Our analysis reveals: probes generalize across various uncertainty sources and datasets, and optimal signals emerge from intermediate rather than final layers. This establishes a principled framework for building deployment-ready AI systems by detecting confidence signals from internal states rather than unreliable outputs.

Figures (3)

• Figure 1: Real-world applications such as assistive AR glasses often encounter low-resolution, corrupted images (middle) rather than the high-quality images (left) typical in laboratory settings. To enable reliable deployment, it is crucial to estimate VLMs' uncertainty and abstain from generating incorrect responses. We propose Latent Representation Probing (LRP) for effective and efficient confidence estimation in VLMs. Our analysis reveals that the most effective probing approaches are: (1) aggregating attention patterns on image token positions across all layers (blue probe, right), and (2) ensembling hidden states from layers that exhibit the strongest confidence signals (red probe, right) through majority vote on output token positions.
• Figure 2: Abstention accuracy evaluated by probes trained on HierText. Baseline represents probes' abstention accuracy on original HierText dataset. LRP demonstrates high generalization capability to various unconfident scenarios, indicating it learns general confidence representation.
• Figure 3: The abstention accuracy of probes trained on each layer's hidden states across four datasets. For SeedBench, HierText and EgoTextVQA, the abstention accuracy first increases and then decreases, indicating that confidence signals are richer in the intermediate instead of final layers.