SMART-Med treats collaboration as a learnable decision instead of a default behavior. The key insight is that not every medical query needs costly multi‑agent deliberation; most can be answered by a single well‑trained agent, while only uncertain cases benefit from a small set of expert agents. Accordingly, an uncertainty‑driven routing mechanism decides adaptively whether to answer directly or to invoke expert collaboration, turning dense multi‑agent reasoning into sparse, query‑aware coordination. To make routing reliable, the authors built the MedVR-3K dataset with high‑quality reasoning traces and a difficulty hierarchy, then fine‑tuned local agents with GRPO using two complementary rewards: a certainty‑aware accuracy reward and a vision‑language reward model for correctness and reasoning quality. Experiments on Slake, PathVQA, VQA‑RAD, and PMC‑VQA show that SMART‑Med matches existing multi‑agent approaches while cutting inference time by more than 3.8× and token usage by 72%. After fine‑tuning on MedVR‑3K, overall performance improves by about 6.7%.
Review