As vision-language models (VLMs) are increasingly applied to medical AI, existing benchmarks mainly focus on evaluating their diagnostic ability over given medical images and texts, implicitly assuming that standardized medical images, texts, or question–answer pairs are already prepared. However, this assumption does not hold when we apply VLMs in real clinical practice, where medical data is often raw, heterogeneous, and fragmented across different sources. In this paper, we study this missing step, i.e., raw medical data standardization. Specifically, models are given raw dataset folders and evaluated on their ability to identify source formats, convert raw medical images into VLM-compatible visual inputs, extract relevant textual information, and organize the results into structured image-text pairs. To construct this Medical Data Standardization Benchmark (MDS-Bench), we manually annotate 1,939 raw medical data standardization tasks covering diverse clinical practice, radiology modalities, annotation formats, and directory layouts. Extensive experiments show that even the best performing VLM, i.e., Gemini 3 Flash, achieves only a 48.6% end-to-end success rate. Our research highlights raw medical data standardization as a critical bottleneck for medical AI diagnosis in real practice.

Predicting next surgical phase, viewed as a form of surgical procedural intent, with high accuracy is essential for robotic real-time surgical systems. Previous approaches struggle with short-term dynamics, occlusions, and phase persistence; end-to-end large vision-language models (LVLMs) can reason from visual inputs but often fail to exploit structured temporal observations. Thus, we introduce SurgAgents, a multi-agent framework for embodied surgical intent prediction. SurgAgents decomposes prediction into specialized agents: (1) a ResNet-based single-frame classifier agent extracts per-frame phase information; (2) a YOLO-based instrument specialist agent extracts per-frame instrument sets; (3) a history descriptor agent composes these facts into frame-level descriptions; (4) a history summarizer agent compresses long histories into a compact temporal narrative; (5) a phase predictor agent assembles the narrative and the current frame into a formatted LVLM prompt; (6) a retrieval agent with a confidence-gated Retrieval-augmented generation (RAG) module selectively retrieves similar case patterns when prediction confidence is low. On Cataract-1K, SurgAgents attains 46.54% macro-F1 with GPT-4o, outperforming the baseline, with systematic ablation studies showing that each agent materially contributes to performance. Latency measurements demonstrate feasibility for robotic real-time surgical support during surgery procedures.