Benchmarking Corruption Robustness of LVLMs: A Discriminative Benchmark and Robustness Alignment Metric

TL;DR

This paper introduces Bench-C, a discriminative benchmark and Robustness Alignment Score to evaluate and analyze the corruption robustness of large vision-language models, revealing nuanced failure and recovery patterns.

Contribution

The paper proposes a new benchmark and metric for assessing corruption robustness in LVLMs, addressing limitations of existing evaluation methods.

Findings

Models show distinct behaviors under corruptions, such as confidence errors.

Subtle corruptions can slightly improve accuracy but degrade prediction structure.

Decomposing robustness reveals different failure and recovery patterns.

Abstract

Despite the remarkable reasoning abilities of large vision-language models (LVLMs), their robustness under visual corruptions remains insufficiently studied. Existing evaluation paradigms exhibit two major limitations: 1) the dominance of low-discriminative samples in current datasets masks the real robustness gap between models; and 2) conventional accuracy-based metric fail to capture the degradation of the underlying prediction structure. To bridge these gaps, we introduce Bench-C, a comprehensive benchmark emphasizing discriminative samples for assessing corruption robustness, where a selection strategy is proposed to jointly consider the prediction inconsistency under corruption and the semantic diversity. Furthermore, we propose the Robustness Alignment Score (RAS), a unified metric that measures degradation in logit-level prediction structure by considering the shifts in prediction uncertainty and calibration alignment. Comprehensive experiments and analysis reveal several interesting findings: 1) model behaviors exhibit distinguish patterns under corruptions, such as erroneous confidence and hesitation; 2) despite subtle corruption may lead to a slight accuracy gain, the overall prediction structure still degrades; 3) by decomposing corruption robustness into destructive and corrective components, the distinct failure and recovery patterns across models can be revealed.