We present RadA-BenchPlat, a two-layer benchmark for radiology agents: a synthetic layer with 2.2k clinician-verified records, 24.2k QA pairs and 10 tool categories, and a real-environment layer pairing 165 2D/3D cases with executable tools. High completion rates under idealized settings do not carry over to real execution, while prompting strategies and on-the-fly synthesis of missing tools substantially narrow the gap.

We introduce MedSR-Copilot, a PRISMA-aligned multi-agent copilot that automates medical systematic reviews end-to-end—literature retrieval, coarse-to-fine screening, data extraction, risk-of-bias assessment and evidence synthesis—together with MedSR-Bench for end-to-end evaluation. It attains 63.6% human-aligned conclusions, and as a copilot cuts review time by 64.9% while raising conclusion accuracy by 27.4 points.

We present OpenPheno, a multimodal foundation model that reframes bioactivity prediction as open-set visual-language question answering. Pairing chemical structures and Cell Painting phenotypic profiles with natural-language assay descriptions enables a "profile once, predict many" paradigm, reaching 0.75 mean AUROC zero-shot on 54 unseen assays and holding 0.66 when both compounds and assays are novel.

We train LLMs as interactive diagnostic agents with end-to-end multi-turn reinforcement learning. DiagGym, a diagnostics world model built from electronic health records, serves as a virtual clinical environment for closed-loop training, and DiagBench supplies 2.2K physician-validated cases with 3.3K rubrics. The resulting DiagAgent outperforms 11 state-of-the-art LLMs and 2 prompt-engineered agents in both in-domain and out-of-domain settings.

We introduce Deep-DxSearch, an end-to-end agentic RAG system trained with reinforcement learning for traceable diagnostic reasoning. Our framework addresses knowledge gaps and hallucinations in medical LLMs by constructing a large-scale medical retrieval corpus and using the LLM as a core agent with tailored rewards. Deep-DxSearch consistently outperforms prompt-engineering and training-free RAG approaches, achieving substantial gains over GPT-4o, DeepSeek-R1, and other medical frameworks for both common and rare disease diagnosis.

We present M³Builder, an LLM-powered multi-agent system for autonomous end-to-end medical imaging AI model training—the first to automate ML in medical imaging, lowering the threshold for clinicians to develop and apply AI models, and promoting the widespread adoption of AI tools in real clinical scenarios.

We build up an academically accessible, large-scale diagnostic dataset, present a knowledge enhanced model architecture that enables processing arbitrary number of input scans from various imaging modalities, and initialize a new benchmark for multi-modal multi-anatomy long-tailed diagnosis. Our method shows superior results and serves as a pre-trained model that can be finetuned to benefit diagnosis on various external datasets.

We present recent efforts on assessing GPT-4V for multimodal medical diagnosis through case studies, covering 17 human body systems across 8 clinical imaging modalities, including radiology and pathology.