A shared visual tokenizer for heterogeneous 2D medical images and 3D volumes. Slice differential regularization preserves anatomical continuity, while multi-expert representation alignment brings complementary clinical semantics into a unified token space. The tokenizer supports classification, retrieval, segmentation, synthesis, visual question answering, and medical report generation.

A benchmark examining medical multimodal models under image quality degradations. It combines modality-specific corruption severity, fine-grained capability assessment, and confidence calibration to study robustness beyond performance on clean images.

BiasRecon studies MRI reconstruction under shifts in imaging centers, anatomy, and acquisition protocols. Bias-calibrated adaptation adjusts a pretrained generative prior using measurement information at test time.

A data engine for improving medical image quality assessment through an evaluate–explore–evolve loop. Model failure prototypes guide retrieval and human-in-the-loop annotation, followed by fine-tuning and renewed evaluation.

A multimodal reinforcement learning framework that assigns visual alignment supervision according to the observed effect of image interventions. PatchSwap interventions and a frozen audit policy route samples toward sensitivity or invariance objectives for visually grounded reasoning.

A unified medical model for multimodal understanding and generation within one set of weights. Progressive training connects image interpretation with interleaved visual and textual outputs, including cross-modal synthesis and image enhancement.

A continual medical segmentation framework that replays jointly generated image-mask pairs to reduce forgetting across changing modalities and tasks. Bayesian Joint Diffusion preserves image-mask correspondence, while shared task-aware prompts coordinate the generator and segmentor.

Radiology report generation that integrates anatomical region features and detected pathologies into report decoding. Regional prompts connect localized image evidence with the resulting report.

A benchmark for evaluating deep evidence integration in medical guideline development. MedProbe-Eval combines task-adaptive quality rubrics with atomic-claim evidence verification to assess language models and deep research agents against expert guideline references.

A benchmark for integrating multimodal and multi-image evidence in rare-disease cases. Four tracks evaluate diagnosis, treatment planning, cross-image evidence alignment, and examination suggestions using evidence-grounded case questions.

A survey and organized catalog of open-access medical imaging datasets for foundation model development. It maps available data resources and the challenges of assembling heterogeneous medical imaging collections.

Implicit Vector Quantization regularizes visual representations to counter low-rank collapse in concept bottleneck models. Magnet Attention aggregates patch features into interpretable concept prototypes, supporting many-to-many relationships between visual features and concepts.

A medical image quality benchmark with complementary perception and descriptive reasoning tasks. It evaluates multimodal models across imaging modalities and compares model-based judgments with expert assessments.

A framework for tracing the historical development of scientific ideas from literature. It combines proposed evolutionary links with citation-grounded verification to support scientific reasoning and assessment.

SGI-Bench evaluates scientific capabilities through tasks aligned with scientists' workflows. The benchmark examines model behavior in scientific problem solving beyond isolated knowledge questions.

Medical image grounding trained with rewards for spatial accuracy and semantic consistency. A Chain-of-Box representation brings referring regions into the model's intermediate reasoning.

Caption-guided retinal image synthesis with fine-grained descriptions of anatomy and pathology. A captioned fundus photograph dataset supports semantic control beyond disease labels and the study of synthetic data for downstream tasks.

Cross-sequence MRI translation formulated as evidential regression. The framework combines uncertainty estimation, multi-source fusion, and distribution calibration to address shifts between medical centers.

Text-guided generation of ophthalmic surgical videos from natural language instructions. A curated video-instruction dataset and progressive tuning adapt a video generation model to ophthalmic workflows.

A data-centered survey of scientific language models, training corpora, and evaluation. It examines the transition from static assessments toward scientific workflows and discusses agents that generate, test, and revise scientific knowledge.

A scientific multimodal foundation model combining general understanding and reasoning with specialist scientific data modalities. This large collaboration develops an open model for tasks across several scientific disciplines.

Unpaired translation between chest X-rays with and without lung opacities. Adaptive activation masks and cross-domain alignment focus modifications on opacity regions, with evaluation of lung segmentation and lesion classification.

MedITok learns a shared token space for medical image reconstruction, synthesis, and interpretation. Its two-stage alignment approach uses unpaired images for structural representation and image-text pairs for clinical semantics.

An investigation of latent diffusion representations for lung disease classification in chest X-rays. The study connects image generation representations with diagnostic classification.

A multimodal medical reasoning model trained with reinforcement learning and synthesized reasoning examples. The work studies medical image diagnosis and visual question answering.

CT airway segmentation with parallel supervised and unsupervised learning. Dynamic masking and multi-order response targets address sparse annotations and airway texture variation.

A cyclic vision-language module explains radiology report generators through counterfactual chest X-ray manipulation. Comparing original and manipulated images identifies visual features associated with changes in generated reports.

Medical image generation conditioned on clinical descriptions and segmentation masks. The work uses vision-language representations to investigate how clinical attributes relate to patterns in generated chest CT images.