Research map / Vision-language and multimodal / Visual question answering
CMID: Towards Medical Visual Question Answering via Contrastive Mutual Information Decoding
Zhihong Zhu, Yunyan Zhang, Fan Zhang, Bowen Xing, Xian Wu
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add attention shift correction (alpha term, focus-excluding contrast) (ablation) | CMID · VQA-RAD open | Recall: 36.57 -> 38.75 | Improved |
| add attention dispersion correction (beta term, focus-preserving contrast) (ablation) | CMID · VQA-RAD open | Recall: 37.12 -> 38.75 | Improved |
| add reliability gate R_PFA instead of fixed 1 (ablation) | CMID · VQA-RAD open | Recall: 36.89 -> 38.75 | Improved |
| use entropy-aware layer reweighting instead of naive averaging (ablation) | CMID · VQA-RAD open | Recall: 37.41 -> 38.75 | Improved |
| use contrastive deltas rather than directly injecting focus-view logits (ablation) | CMID · PathVQA open | Recall: 10.61 -> 11.41 | Improved |
| hard binary mask instead of soft mask (ablation) | CMID PFA · VQA-RAD avg | Score: 52.15 -> 52.81 | Improved |
| aggregate attention from all layers instead of a layer subset (ablation) | CMID PFA · VQA-RAD avg | Score: 51.34 (early) / 52.56 (middle) -> 52.81 (all) | Improved |
| apply CMID to other medical LVLMs | HuatuoGPT-V, Med-Flamingo, RadFM, MedVInT, LLaVA-Med · VQA-RAD, SLAKE | Score: 48.08 -> 52.81 (LLaVA-Med VQA-RAD); 44.82 -> 47.26 (RadFM) | Improved |
| general-domain contrastive decoding (M3ID, VCD) on Med-VQA | LLaVA-Med v1.5 · VQA-RAD open | Recall: 34.39 -> 30.93 (M3ID); 34.39 -> 29.75 (VCD) | Hurt |
| increase alpha or beta beyond optimum (ablation) | CMID · PathVQA | Overall Score | Hurt |
| add CMID to greedy decoding | LLaVA-Med v1.5 · Inference cost | latency ms/token: 121.48 -> 136.23 | Hurt |
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.
- SparseVLM — SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inference (ICML 2025) · better on 1 of 1 shared result
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
- Falcon Medical Visual Question Answering (AAAI 2025 · similarity 0.92)
- Fine-grained Adaptive Visual Prompt for Generative Medical Visual Question Answering (AAAI 2025 · similarity 0.91)
- Alignment, Mining and Fusion: Representation Alignment with Hard Negative Mining and Selective Knowledge Fusion for Medical Visual Question Answering (CVPR 2025 · similarity 0.91)
- DiN: Diffusion Model for Robust Medical VQA with Semantic Noisy Labels (CVPR 2025 · similarity 0.90)
- Visual and Domain Knowledge for Professional-level Graph-of-Thought Medical Reasoning (ICML 2025 · similarity 0.90)
- OmniMedVQA: A New Large-Scale Comprehensive Evaluation Benchmark for Medical LVLM (CVPR 2024 · similarity 0.90)
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.
