VAUQ: Vision-Aware Uncertainty Quantification for LVLM Self-Evaluation

TL;DR

VAUQ is a novel framework for LVLM self-evaluation that explicitly measures visual evidence influence to improve correctness estimation, outperforming existing methods.

Contribution

It introduces the Image-Information Score and a core-region masking strategy for vision-aware uncertainty quantification in LVLMs.

Findings

VAUQ outperforms existing self-evaluation methods across multiple datasets.

The Image-Information Score effectively captures visual input's impact on predictions.

Combining entropy with IS provides a reliable, training-free correctness score.

Abstract

Large Vision-Language Models (LVLMs) frequently hallucinate, limiting their safe deployment in real-world applications. Existing LLM self-evaluation methods rely on a model's ability to estimate the correctness of its own outputs, which can improve deployment reliability; however, they depend heavily on language priors and are therefore ill-suited for evaluating vision-conditioned predictions. We propose VAUQ, a vision-aware uncertainty quantification framework for LVLM self-evaluation that explicitly measures how strongly a model's output depends on visual evidence. VAUQ introduces the Image-Information Score (IS), which captures the reduction in predictive uncertainty attributable to visual input, and an unsupervised core-region masking strategy that amplifies the influence of salient regions. Combining predictive entropy with this core-masked IS yields a training-free scoring function that reliably reflects answer correctness. Comprehensive experiments show that VAUQ consistently outperforms existing self-evaluation methods across multiple datasets.