Research map / Vision-language and multimodal / Multimodal LLMs

How Do Medical MLLMs Fail? A Study on Visual Grounding in Medical Images

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Benchmarks and analysis of MLLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Systematically diagnoses weak visual grounding in eight medical MLLMs using a new attention-based benchmark (VGMED) and proposes VGRefine, a training-free attention-knockout method that improves HuatuoGPT-Vision on six Med-VQA benchmarks.

Problem

Medical MLLMs underperform in zero-shot medical VQA, and it is unclear whether this stems from semantic grounding (lack of medical knowledge) or visual grounding (failure to localize relevant image regions).

Approach

Measure visual grounding by comparing attention maps against expert-annotated bounding boxes using AR, KL and JS metrics on a clinically co-designed VQA set, then refine attention at inference time by selecting top-K grounding-relevant heads (chosen on COCO) and knocking out attention to low-activation visual tokens at a single layer.

Key results

• VGRefine on HuatuoGPT-V-7B raises the weighted average on four closed-ended Med-VQA benchmarks from 65.3% to 68.4%, with +11.3 on PathVQA.
• On OmniMedVQA, VGRefine improves average accuracy from 71.3% to 74.4% across eight imaging modalities.
• Clinicians preferred VGRefine attention maps in 76% of 20 blinded cases.

Surprising findings

• The general-domain LLaVA-v1.5 also shows weak attention alignment on medical images, while medical MLLMs ground well on natural images.
• Replacing the CLIP encoder with BiomedCLIP in HuatuoGPT-Vision did not solve the visual grounding issue.

Limitations

• Visual grounding is measured via attention maps, which are an imperfect proxy for grounding.
• Evaluation of the VGMED questions relies on bounding boxes converted from segmentation masks and on GPT-4-generated questions.
• Gains on the zero-shot benchmark are evaluated on only the closed-ended subset in the main text.

Experimental setup

Datasets: VGMED MS COCO VQA-RAD SLAKE PathVQA PMC-VQA OmniMedVQA MMMU

Benchmarks: VQA-RAD SLAKE PathVQA PMC-VQA OmniMedVQA MMMU (Health & Medicine)

Metrics: attention ratio KL divergence JS divergence accuracy BLEU-1 BERT score OpenRecall human preference rate

Baselines: Qwen-VL-Chat LLaVA-v1.6-7B Med-Flamingo RadFM LLaVA-Med-7B LLaVA-Tri VILA-M3-8B MedPLIB PAI AdaptVis ViCrop

Builds on: LLaVA-Med HuatuoGPT-Vision LLaVA-v1.5 BiomedCLIP

Closest papers by meaning

• MedSG-Bench: A Benchmark for Medical Image Sequences Grounding (NeurIPS 2025 · similarity 0.93)
• MedMO: Grounding and Understanding Multimodal Large Language Model for Medical Images (CVPR 2026 · similarity 0.91)
• GMAI-MMBench: A Comprehensive Multimodal Evaluation Benchmark Towards General Medical AI (NeurIPS 2024 · similarity 0.90)
• MedSIGHT: Towards Grounded Visual Comprehension in Medical Large Vision-Language Models (ICML 2026 · similarity 0.90)
• OmniMedVQA: A New Large-Scale Comprehensive Evaluation Benchmark for Medical LVLM (CVPR 2024 · similarity 0.90)
• Interpretable Bilingual Multimodal Large Language Model for Diverse Biomedical Tasks (ICLR 2025 · similarity 0.90)