Large language models (LLMs) have shown promise in medical question answering and clinical reasoning, yet their progress is limited by static parametric knowledge and expensive expert supervision. Self‑evolving agents address this by iteratively generating tasks and solving them, offering a scalable improvement path. Most existing self‑evolving approaches target domains like mathematics or coding where solutions can be verified exactly via answers or executable code. Medical reasoning, however, is open‑ended, knowledge‑intensive, and often only partially verifiable.\ \ We introduce MedZERO, a self‑evolving framework tailored for open‑ended medical reasoning. MedZERO consists of two core components:\
- Examiner: generates frontier medical question‑option pairs to create new learning tasks;\
- Reasoner: solves these tasks through evidence‑grounded multi‑turn reasoning, leveraging external knowledge tools such as retrieval systems.\ \ To ensure reliable continual improvement, MedZERO adopts controlled knowledge accumulation, separating temporary exploratory knowledge from curated persistent knowledge during reasoning. This design preserves a space for novel information while preventing noisy data from contaminating long‑term memory.\ \ We evaluate MedZERO on five public medical reasoning benchmarks using 4B‑ and 8B‑scale base models under open‑ended evaluation. Across all settings, MedZERO consistently outperforms the underlying base models and prior self‑evolving baselines, achieving up to a 13.7‑point average accuracy gain over the next‑best baseline.\ \ Review