A Vision-Language Benchmark for Microscopy Understanding, featuring 17,000 microscopy images sourced from 24 publicly available datasets. As the most diverse microscopy benchmark to date, it spans light (LM), fluorescence (FM), and electron microscopy (EM) across 8 sub-modalities, 91 distinct cell, tissue, and structure types, and 24 staining techniques. Microbench supports tasks including closed-form visual question answering (VQA), object detection, and segmentation.

We introduce MedAlign, a benchmark of 983 natural language instructions about EHR data. MedAlign is curated by 15 clinicians (7 specialities), includes human-written reference responses, and provides 276 full longitudinal EHRs for grounding instruction-response pairs.

Introducing Clinfo.ai, the first open-source agentic system designed to answer medical questions using scientific literature. Clinfo.ai employs a chain of large language models to convert a question into a query and explore the most relevant literature to provide an up-to-date answer.

Unsupervised shape representations of cells and organelles are often erroneously sensitive to image orientation. We introduce O2VAE, an orientation-invariant autoencoder that mitigates this issue by using equivariant convolutional network encoders.