Modality robustness under knowledge conflict is studied across 13 MLLMs and two datasets, and it is found that instability has practical consequences: it degrades performance in multimodal RAG and can be exploited by adversarial attacks.
Abstract
Multimodal large language models (MLLMs) are increasingly provided with contextual evidence in heterogeneous forms: as a text passage, as a rendered image of the same passage, or as both together. However, it remains unclear how consistently these surface forms are processed, especially when the evidence conflicts with the model's parametric knowledge. We study modality robustness under knowledge conflict across 13 MLLMs and two datasets, and find them far from robust. (1) Contrary to common belief, models favor a context that contradicts parametric knowledge more readily in image form than in text form; (2) when a contradicting text and image are presented together, the preferred modality is essentially arbitrary, varying with input order, model, and dataset. We further demonstrate that this instability has practical consequences: it degrades performance in multimodal RAG and can be exploited by adversarial attacks. To alleviate this brittleness, we examine several simple techniques---prompting, steering, supervised fine-tuning (SFT), and direct preference optimization; the majority prove ineffective, whereas SFT achieves moderate success. We therefore call for greater awareness of this inconsistency and argue that it is fundamental, demanding attention at multiple training stages.
A 998-case diagnostic is introduced that independently varies visual evidence, commonsense priors, and external text, and probes when this failure arises by moving the information boundary around a context-blind visual witness.
Dual-encoder vision-language models (VLMs) allow to embed texts and images in a shared representation space. However, it has been shown that these models are subject to a “modality gap phenomenon” meaning there exists a clear separation between the embeddings from one modality and another in the embedding space. While...
François Role, Sébastien Meyer, Victor Amblard· International Journal of Mul...· 0 citations
This paper proposes a novel phase-aware adversarial attack framework that explicitly restricts adversarial perturbations to structure-relevant phase regions to suppress the MLLM' focus for effective and imperceptible attacks.
Dai-Zong Liu, Junhao Dong, Zhi-Yuan Ma et al.· IEEE transactions on multime...· 1 citation
Adaptive Relevance-guided Evidence Allocation (AREA) is introduced, a training-free inference-time method that formulates evidence highlighting as adaptive allocation and establishes the best performance among training-free highlighting methods.
Zhen-Bin Wang, Lei Zhang, Li-Tuan Wang et al.· 0 citations
ForUM is presented, a training-free test-time fusion of frozen MLLMs guided by two fixed geometric rules: agreement-based selection keeps the region supported by the most distinct models, and medoid localization returns an actual member box instead of a coordinate average, so one loose prediction cannot shift the answe...
Adversarial Contrastive Fine-Tuning (ACFT) uses an Adversarial Hallucination Attribute Flipping procedure, involving minimal, targeted adversarial perturbations that flip an image's hallucination attribute, to construct perfectly aligned positive-negative pairs, which are then used for contrastive fine- tuning.
Pei-Yang Xu, Xiaopei Zhu, Jun Zhu et al.· 0 citations
Related blog posts
MIT News · Artificial Intelligence· news.mit.eduNov 19, 2026
An MIT and Dana-Farber Institute-led team used large language models to examine nearly 4,000 studies and hundreds of code repositories, revealing gaps in the information needed to scrutinize and reuse published research.
By studying the computational foundations of language learning, Associate Professor Jacob Andreas builds intelligent systems that can communicate effectively with humans.