AI Paper Atlas

Research map / Vision-language and multimodal / Visual question answering

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

Beidi Zhao, Wenlong Deng, Xinting Liao, Yushu Li, Nazim Shaikh, Yao Nie, Xiaoxiao Li

ICML 2026 · Vision-language and multimodal › Visual question answering · approach: Multimodal LLM question answering · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add MAD-RAG (dual question plus attention mixing) to vanilla RAGVanilla RAG · E-VQA, LLaVA-1.5-7Baccuracy (%): 53.89 -> 63.09Improved
add attention mixing on top of dual-question input (ablation)MAD-RAG · OK-VQA, LLaVA-1.5-7Baccuracy (%): 69.65 -> 70.22Improved
add attention mixing on top of dual-question input (ablation)MAD-RAG · OK-VQA, Qwen2.5-VL-7Baccuracy (%): 63.85 -> 64.91Improved
duplicate the question (dual-question) without attention intervention (ablation)Vanilla RAG · OK-VQA, LLaVA-1.5-7Baccuracy (%): 65.46 -> 69.65Improved
duplicate the question (dual-question) without attention intervention (ablation)Vanilla RAG · OK-VQA, Qwen2.5-VL-7Baccuracy (%): 63.80 -> 63.85No clear effect
place question after the retrieved context (Swap Q/C) (ablation)Vanilla RAG · OK-VQA, LLaVA-1.5-7Baccuracy (%): 65.46 -> 68.17Improved
place question after the retrieved context (Swap Q/C) (ablation)Vanilla RAG · OK-VQA, Qwen2.5-VL-7Baccuracy (%): 63.80 -> 61.92Hurt
add retrieved context (RAG) versus closed-bookLLaVA-1.5-7B · image attention ratioimage attention ratio: 15.77% -> 9.07%Hurt
apply attention mixing weight alpha between 0.1 and 0.5 (ablation)MAD-RAG · OK-VQA, E-VQA, InfoSeek, LLaVA-1.5-7BaccuracyNo clear effect
use MAD-RAG instead of ALFAR on Qwen2.5-VL-7B OK-VQAALFAR · OK-VQA, Qwen2.5-VL-7Baccuracy (%): 63.55 -> 64.91Improved
add VLM hallucination mitigation methods (VHR) to RAGVanilla RAG · E-VQA, Qwen2.5-VL-7Baccuracy (%): 82.21 -> 58.67Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy