We release MedQ-Deg, a multidimensional benchmark evaluating 40 mainstream MLLMs across 18 degradation types, 30 capability dimensions, and 7 imaging modalities (24,894 QA pairs). We reveal the AI Dunning-Kruger Effect — models maintain high confidence despite severe accuracy collapse under degraded inputs.

We introduce UniMedVL, the first medical multimodal model that unifies image understanding and generation in a single architecture. Built on 5.6M multimodal samples and Progressive Curriculum Learning, UniMedVL achieves strong results on 5 understanding benchmarks and 8 generation modalities.

We release MedQ-Bench, a novel perception-reasoning benchmark for medical image quality assessment covering 5 modalities and 40+ quality attributes (3,308 samples). Zero-shot evaluation of 14 leading MLLMs shows GPT-4 achieves 68.97% accuracy — still 13.5% below expert performance.

We welcome Wenhao Tang (Nankai University), Shujian Gao (Fudan University), and Jiashi Lin (Northwestern Polytechnical University) as new visiting researchers, bringing expertise in computational pathology, multimodal learning, and LLM-based agents.

We release a comprehensive survey of Scientific Large Language Models (Sci-LLMs), produced in collaboration with 20+ leading global institutions. The survey covers 1,000+ papers and 600+ key datasets, and proposes a roadmap for AI-assisted scientific discovery ecosystems.

OphCLIP, our hierarchical retrieval-augmented framework for ophthalmic surgical video understanding, was accepted at ICCV 2025. Built on OphVL (375K+ video-text pairs), OphCLIP sets new records on 11 benchmarks for phase recognition and multi-instrument detection tasks.

Our text-guided ophthalmic surgical video generation model Ophora was accepted at MICCAI 2025 as an oral presentation. Ophora is trained on 160K video-text pairs and outperforms existing methods on FID, FVD, and CLIPScore metrics.

Three papers accepted at MICCAI 2025: RetinaLogos (language-driven high-resolution fundus image generation), ProgEmu (interpretable counterfactual medical image generation), and Multi-modal MRI Translation via Evidential Regression and Distribution Calibration.

We release MedITok, the first unified visual tokenizer designed for medical images. Pre-trained on 30M+ images, MedITok achieves SOTA across reconstruction, classification, generation, and VQA tasks spanning 9 imaging modalities and 30+ datasets.

MedSegAgent, our multi-agent system for instructive medical image segmentation, was accepted at IEEE Journal of Biomedical and Health Informatics (JBHI). The system supports 343 segmentation targets across CT, MRI, PET/CT, and ultrasound without training a single universal model.

Our whole-slide pathology image understanding assistant SlideChat was accepted at CVPR 2025. SlideChat achieves 81.17% accuracy on SlideBench-VQA and surpasses state-of-the-art on 18 of 22 benchmark tasks.

GMAI-MMBench was presented at NeurIPS 2024 — the most comprehensive general medical AI evaluation platform covering 284 datasets, 38 imaging modalities, and 18 clinical tasks. Even top-performing GPT-4o achieves only 53.96% accuracy.

We release GMAI-VL, a general-purpose medical vision-language model trained on GMAI-VL-5.5M — 5.5M high-quality image-text pairs spanning 18 clinical specialties and 10+ imaging modalities. GMAI-VL achieves or surpasses SOTA on multiple medical multimodal VQA and diagnostic reasoning benchmarks.

SAM-Med3D was selected as an oral presentation at the ECCV 2024 Biomedical Image Computing (BIC) Workshop. SAM-Med3D adapts SAM to 3D volumetric medical images, covering 247 segmentation categories across 21K medical volumes.

OmniMedVQA was accepted at CVPR 2024. The benchmark integrates 73 datasets across 12 imaging modalities and 20+ anatomical regions, revealing that many medical-specific models surprisingly underperform general-purpose LVLMs on medical VQA tasks.