When RAG Hurts: Diagnosing and Mitigating Attention Distraction in Retrieval-Augmented LVLMs

While Retrieval-Augmented Generation (RAG) is one of the dominant paradigms for enhancing Large Vision-Language Models (LVLMs) on knowledge-based VQA tasks, recent work attributes RAG failures to insufficient attention towards the retrieved context, proposing to reduce the attention allocated to image tokens. In this work, we identify a distinct failure mode that previous study overlooked: Attention Distraction (AD). When the retrieved context is sufficient (highly relevant or including the correct answer), the retrieved text suppresses the visual attention globally, and the attention on image tokens shifts away from question-relevant regions. This leads to failures on questions the model could originally answer correctly without the retrieved text. To mitigate this issue, we propose MAD-RAG, a training-free intervention that decouples visual grounding from context integration through a dual-question formulation, combined with attention mixing to preserve image-conditioned evidence. Extensive experiments on OK-VQA, E-VQA, and InfoSeek demonstrate that MAD-RAG consistently outperforms existing baselines across different model families, yielding absolute gains of up to 4.76%, 9.20%, and 6.18% over the vanilla RAG baseline. Notably, MAD-RAG rectifies up to 74.68% of failure cases with negligible computational overhead.

Retrieval-augmented large vision–language models (LVLMs) improve visual question answering by using external knowledge.However, even when the model can answer a question correctly on its parametric knowledge, adding high-quality oracle retrieved context can cause it to fail. We identify a failure mode called attention distraction (AD), including (1) cross-modal AD: retrieved context suppresses attention allocated to image tokens, and (2) intra-image AD: visual attention shifts away from question-relevant visual regions to irrelavant regions. To address this, we propose MAD-RAG, a training-free method that separates visual grounding from knowledge integration and recover the visual grounding ability of retrieval-augmented LVLMs. The finding of AD and MAD-RAG leads to more accurate and reliable answers across knowledge-based VQA benchmarks, with minimal additional computational cost, making it practical for real-world multimodal systems.