Open Ended Medical Reinforcement Learning
Synthetic medical VQA pipeline: 119K images annotated by frontier VLMs, cross-validated at 93% agreement, fine-tuned on 3 model families (2-3B params)
[ICMR'21, Best Poster Paper Award] Medical Visual Question Answering with Multi-task Pre-training and Cross-modal Self-attention
Official repository for the ACL 2025 Findings paper "Worse than Random? An Embarrassingly Simple Probing Evaluation of Large Multimodal Models in Medical VQA"
(ECCV 2026) Official repository for MedQ-Deg, a benchmark evaluating medical VLM robustness to image degradation. Features 24,894 QA pairs across 7 imaging modalities, 18 degradation types, and 30 clinical capabilities, with multi-severity evaluation and confidence calibration analysis.
[IEEE TMI'22] VQAMix: Conditional Triplet Mixup for Medical Visual Question Answering
Research framework evaluating retrieval-augmented vision-language models for evidence-grounded medical visual question answering.
Visual Question Answering in the Medical Domain
Visual Grounding Score Guided Decoding for Hallucination Mitigation in Medical VLMs
Single-GPU GRPO post-training for medical visual diagnostic reasoning. Qwen3-VL-4B + LoRA, verifiable rewards, no reasoning annotations. DAPO-style dynamic sampling cuts zero-gradient steps from 137/200 to 25/200.
MRareBench is a multimodal benchmark for evaluating rare-disease diagnosis and patient-specific visual evidence grounding across diverse medical imaging modalities.
A curated collection of gastrointestinal endoscopy datasets for AI research, covering colonoscopy, polyp detection, segmentation, classification, capsule endoscopy, surgical endoscopy, medical VQA, and multimodal learning.
A curated literature resource hub for Medical Visual Question Answering, covering surveys, datasets, benchmarks, evaluation metrics, representative methods, and multimodal medical agents, with a focus on the shift from passive answer prediction to active, evidence-seeking clinical inquiry.
Training-free audits of output-schema anchoring and label reassignment in medical multimodal MCQ evaluation.
AstraQ-VL is a minimal, efficient astronomy Vision-Language Model following the LLaVA architecture. Stage 1 trains only a lightweight MLP connector to align frozen CLIP vision features with a frozen LLM. Stage 2 then warm-starts that connector and fine-tunes the LLM with LoRA adapters on instruction (QA) data.
When will a VLM fail on medical VQA? Image-free risk prediction and shared error structure (ICASSP 2027 submission): code, splits, prompts, pre-registration, result tables, generations
Agentic multimodal radiology question answering with image-grounded verification
基于 Qwen2.5-VL-7B 的皮肤病变视觉问答系统：LoRA 微调 + vLLM 部署 + FastAPI + 微信小程序
Official implementation for LiteMedRAG: Selective Retrieval-Augmented Compact Medical Visual Question Answering (MIWAI 2026)
Official implementation of our MICCAI MedReason 2026 medical VQA system with option-aware retrieval and task-specific VLM adaptation.
To associate your repository with the medical-vqa topic, visit your repo's landing page and select "manage topics."