PRISM: A Benchmark for Unveiling Cross-modal Knowledge Inconsistency in Large Vision-Language Models

摘要

Recent advances in Large Vision-Language Models (LVLMs) have unearthed boosted performance of multi-modal understanding. In this paper, however, we for the first time uncover a critically under-explored challenge persisting in this trend, that LVLMs unfortunately exhibit cross-modal knowledge inconsistencies. Cross-modal knowledge inconsistency refers to the tendency of providing semantically inconsistent responses to contexts that are semantically equivalent but expressed in different modalities. In real-world applications, users can rely on either text or image to express their ideas. Inconsistent responses across modalities can confuse users, challenging the reliabilities of LVLMs in practice. Therefore, we argue that evaluating performance on either multi-modal or text-only task is insufficient; and waiving the mentioned cross-modal knowledge inconsistency is crucial. The paper proposes PRISM, the first-ever benchmark for measuring the inconsistency, and the corresponding evaluation metric Know-Inc. PRISM covers commonsense, encyclopedia, and mathematics knowledge, with manually-screened samples of semantic alignment. From the evaluation results of up to 27 LVLMs with diverse structures, we conclude that: 1) LVLMs show a preference for textual input, 2) there is a correlation between inconsistency and accuracy, and 3) the inconsistency is more prominent in encyclopedia knowledge. These findings can shed light on further optimization and development of LVLMs.