Research map / Vision-language and multimodal / Multimodal LLMs
MEDA: Medical-Oriented Activation Editing for Hallucination Mitigation in Medical Large Vision-Language Model
Tianbo Wang, Yuqing Ma, Lingyan Meng, Zhange Zhang, Kewei Liao, Jian Yang, Simin Li, Jinyang Guo, Xianglong Liu
Read the paperCheck your idea against 137,217 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.
Proposes MEDA, a medical-oriented activation editing method that mitigates hallucinations in medical large vision-language models by capturing imaging manifestations and diagnostic principles.
Problem
Medical large vision-language models (Med-LVLMs) suffer severe hallucinations that pose safety risks in clinical deployment; existing activation-editing methods lack the medical domain expertise needed to capture imaging manifestations and diagnostic principles.
Approach
Adapts activation editing to the medical domain so that the edit directions encode medical imaging manifestations and diagnostic principles, rather than generic visual features, to suppress hallucinations at inference time with minimal cost.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add QMS manifestation steering (ablation) | LLaVA-Med-1.5 · IU-Xray (VQA) | accuracy: 74.2 -> 81.2 (7.0%) | Improved |
| add PDI diagnostic principle steering (ablation) | LLaVA-Med-1.5 · IU-Xray (VQA) | accuracy: 74.2 -> 82.5 (8.3%) | Improved |
| combine QMS and PDI cooperatively (ablation) | LLaVA-Med-1.5 · IU-Xray (VQA) | accuracy: 82.5 -> 84.8 | Improved |
| add threshold control for manifestation retrieval (ablation) | QMS · IU-Xray (VQA) | accuracy: 80.1 -> 81.2 | Improved |
| add priority weight for entity extraction (ablation) | QMS · IU-Xray (VQA) | accuracy: 78.8 -> 81.2 | Improved |
| add prescribed-order sort to local evidence (ablation) | PDI · IU-Xray (VQA) | accuracy: 81.9 -> 82.5 | Improved |
| add scope-conditioned diagnostic template (ablation) | PDI · IU-Xray (VQA) | accuracy: 80.2 -> 82.5 | Improved |
| use only the raw report instead of template and sort (ablation) | PDI · IU-Xray (VQA) | accuracy: 82.5 -> 79.8 | Hurt |
| apply MEDA instead of best editing baseline AFTER | LLaVA-Med-1.5 · IU-Xray (VQA) | accuracy: 81.4 -> 84.8 | Improved |
| increase editing strength and heads beyond K=32, alpha=5 (ablation) | MEDA · IU-Xray (VQA) | accuracy: 84.8 -> 61.6 | Hurt |
| reduce vector-construction samples to 10 (ablation) | MEDA · MIMIC-CXR | accuracy: 78.2 vs 79.4 | No clear effect |
| apply MEDA to five other LVLMs | MEDA · MIMIC-CXR | accuracy: average +5.9% accuracy, +5.4% RaTEScore | Improved |
Later papers that compared with this method
1 later accepted paper reports MEDA as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- DAVID: Dual-stage Adaptive Vision-text Integrated Decoupling for Multimodal KV Cache Eviction (AAAI 2026 · better on 2 of 2 shared results)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- AFTER — AFTER: Mitigating the Object Hallucination of LVLM via Adaptive Factual-Guided Activation Editing (ICLR 2026) · better on 8 of 9 shared results
- MMedRAG — MMed-RAG: Versatile Multimodal RAG System for Medical Vision Language Models (ICLR 2025) · better on 4 of 4 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.
Experimental setup
Builds on: activation editing
Closest papers by meaning
- AFTER: Mitigating the Object Hallucination of LVLM via Adaptive Factual-Guided Activation Editing (ICLR 2026 · similarity 0.89)
- Anatomical Region-Guided Contrastive Decoding: A Plug-and-Play Strategy for Mitigating Hallucinations in Medical VLMs (AAAI 2026 · similarity 0.89)
- Dynamic Multimodal Activation Steering for Hallucination Mitigation in Large Vision-Language Models (ICLR 2026 · similarity 0.88)
- Intervene-All-Paths: Unified Mitigation of LVLM Hallucinations across Alignment Formats (NeurIPS 2025 · similarity 0.87)
- Look Closer! An Adversarial Parametric Editing Framework for Hallucination Mitigation in VLMs (AAAI 2026 · similarity 0.87)
- Steer Where It Matters: Token-Level Visual-Sensitivity Steering for LVLMs Hallucination Mitigation (ICML 2026 · similarity 0.87)
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.
