Does It Fail to See or Fail to Know? Attributing Errors in Vision-Language Models

2026-07-06

AI Summary

This work addresses the challenge that vision-language models often err when relying on external knowledge and struggle to enable fine-grained diagnosis of error sources—such as failures in perception, recognition, or knowledge gaps. The paper proposes the first unified framework that predicts the root cause of potential failures prior to generation by analyzing visual token representations and the hidden states conditioned on prompts, thereby disentangling perception/recognition bottlenecks from knowledge-related errors. Experimental results demonstrate that recognition errors are predominantly captured by visual tokens, whereas knowledge errors rely more heavily on prompt-conditioned hidden states. This distinction confirms the predictability of failure origins and facilitates efficient, targeted intervention strategies.

Abstract

Vision-language models (VLMs) perform well on visual question answering with high-quality images but struggle when questions require knowledge beyond what is clearly and directly visible. In such settings, uncertainty quantification should not only indicate whether the model is likely to fail but also diagnose why it is uncertain, across dimensions such as perception, entity recognition, and knowledge retrieval. While prior work has focused on individual failure modes in isolation or treated incorrect answers as monolithic failures, we propose a unified framework for disentangling these failure modes and investigate whether pre-generation signals can predict these failure sources. Across a range of datasets and model families, we find a consistent pattern in VLM errors: some failures arise from visual or recognition bottlenecks, while others persist after the relevant entity is identified. Our main finding is that these failure sources can be predicted before decoding: recognition-related failures are best captured by visual-token representations, while failures that remain after recognition are better captured by prompt-conditioned hidden states. This pre-generation signal enables efficient failure-source prediction before the model produces an answer, allowing uncertain cases to be routed to targeted interventions such as image repair, entity recognition support, or external retrieval.

Problem

• vision-language models
• failure attribution
• uncertainty quantification
• visual question answering
• error diagnosis

Innovation

• vision-language models
• failure attribution
• uncertainty quantification
• pre-generation signals
• error diagnosis