Research map / Vision-language and multimodal / Visual question answering
Synthesizing High-Quality Visual Question Answering from Medical Documents with Generator-Verifier LMMs
Xiaoke Huang, Ningsen Wang, Hui Liu, Xianfeng Tang, Yuyin Zhou
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 rubric-guided verification to rubric-guided generation (ablation) | MedVLSynther 3B · Six medical VQA benchmarks average | accuracy (avg): 54.72 -> 55.85 | Improved |
| add rubric-guided verification to rubric-guided generation (ablation) | MedVLSynther 7B · Six medical VQA benchmarks average | accuracy (avg): 57.33 -> 57.56 | Improved |
| use rubric-guided context-aware generation instead of PMC-style text-only generation (ablation) | MedVLSynther 7B · Six medical VQA benchmarks average | accuracy (avg): 54.41 -> 57.33 | Improved |
| use rubric-guided generation instead of PMC-style text-only generation (ablation) | MedVLSynther 3B · Six medical VQA benchmarks average | accuracy (avg): 54.80 -> 54.72 | No clear effect |
| increase training data from 1K to 5K items (ablation) | MedVLSynther 3B · Six medical VQA benchmarks average | accuracy (avg): 52.64 -> 55.85 | Improved |
| increase training data beyond 5K to 10K and full (ablation) | MedVLSynther 3B · Six medical VQA benchmarks average | accuracy (avg): 55.85 -> 55.03 (10K), 55.17 (Full) | No clear effect |
| increase training data from 5K to full 13K items (ablation) | MedVLSynther 7B · Six medical VQA benchmarks average | accuracy (avg): 57.56 -> 58.15 | Improved |
| use stronger verifier GLM-4.5V 108B as both generator and verifier (ablation) | MedVLSynther 7B · Six medical VQA benchmarks average | accuracy (avg): 57.56 -> 58.08 | Improved |
| use same GLM-4.5V model as generator and verifier instead of Qwen2.5-VL 72B verifier (ablation) | MedVLSynther 3B · Six medical VQA benchmarks average | accuracy (avg): 55.85 -> 54.59 | Hurt |
| use RL instead of SFT on MedSynVQA (ablation) | Qwen2.5-VL 3B · Six medical VQA benchmarks average | accuracy (avg): 49.11 -> 55.85 | Improved |
| use MedSynVQA instead of text-only m23k data for RL (ablation) | Qwen2.5-VL 7B · Six medical VQA benchmarks average | accuracy (avg): 54.88 -> 57.56 | Improved |
| use MedSynVQA instead of PMC-VQA data for RL (ablation) | Qwen2.5-VL 3B · Six medical VQA benchmarks average | accuracy (avg): 52.28 -> 55.85 | Improved |
Closest papers by meaning
- Falcon Medical Visual Question Answering (AAAI 2025 · similarity 0.92)
- OmniMedVQA: A New Large-Scale Comprehensive Evaluation Benchmark for Medical LVLM (CVPR 2024 · similarity 0.91)
- DiN: Diffusion Model for Robust Medical VQA with Semantic Noisy Labels (CVPR 2025 · similarity 0.91)
- MediConfusion: Can you trust your AI radiologist? Probing the reliability of multimodal medical foundation models (ICLR 2025 · similarity 0.90)
- Fine-grained Adaptive Visual Prompt for Generative Medical Visual Question Answering (AAAI 2025 · similarity 0.90)
- Alignment, Mining and Fusion: Representation Alignment with Hard Negative Mining and Selective Knowledge Fusion for Medical Visual Question Answering (CVPR 2025 · 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.
