AI Paper Atlas

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

MR-RAG: Multimodal Relevance-Aware Retrieval-Augmented Generation for Medical Visual Question Answering

Xuze Li, Haozhao Wang, Zhenyu Huang, Zhongxu Wang, Jinghua Zhang, Ruixuan Li

CVPR 2026 · Vision-language and multimodal › Visual question answering · approach: Multimodal LLM question answering

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
replace single image-to-text retrieval with multimodal cooperative retrieval (MCR) (ablation)LLaVA-Med-1.5 RAG · Average of three medical VQA datasetsF1 (%): 69.61 -> 73.34Improved
add importance-aware information flow augmentation (IFA) on top of MCR (ablation)MR-RAG · Average of three medical VQA datasetsF1 (%): 73.34 -> 75.33Improved
use fused multi-score retrieval instead of single-score similarity (ablation)MCR · Retrieval similarity ablationaccuracy (%): 77.51 -> 81.95 (vs best single I2T)Improved
use text-to-image similarity alone for retrieval (ablation)MCR · Retrieval similarity ablationaccuracy (%): 77.51 -> 64.71 (vs I2T)Hurt
mix in more irrelevant documents (up to half of input)LLaVA-Med-1.5 RAG · Noise analysisaccuracy (%): 77.9 -> 68.1Hurt
block attention flow Reports to Question in shallow-to-mid layers (ablation)LLaVA-Med-1.5 · IU-Xrayaccuracy differenceHurt
block attention flow Reports to Image or Reports to Last (ablation)LLaVA-Med-1.5 · IU-Xrayaccuracy differenceNo clear effect
apply MR-RAG instead of beam search/greedy across retrieval sizes K from 5 to 30MR-RAG · Harvard-FairVLMedaccuracy range across K: beam search 65.2%-73.8% vs MR-RAG 79.5%-82.1%Improved
add MMed-RAG retrieval and preference alignment to LLaVA-Med-1.5LLaVA-Med-1.5 · Harvard-FairVLMedaccuracy (%): 77.35 -> 70.57Hurt
add VCD decoding to LLaVA-Med-1.5LLaVA-Med-1.5 · MIMIC-CXRaccuracy (%): 71.92 -> 65.35Hurt

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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