ResidencyRL: Reinforcement Learning for Clinical AI in Simulated Environments

A new arXiv preprint introduces ResidencyRL, a reinforcement learning (RL) method designed to train clinical artificial intelligence (AI) agents through simulated multi-turn patient encounters. The method addresses a gap in medical AI: while large language models (LLMs) perform well on static benchmarks, optimizing the full sequence of clinical decisions remains underdeveloped. ResidencyRL pairs a policy agent with LLM simulators that can exhibit complex, adversarial behaviors, training against a structured reward aligned with diagnostic accuracy. Each training trajectory can include up to 60 dialogue turns and 8 tool calls. The approach draws an analogy to medical residency, where physicians convert academic knowledge into clinical expertise through years of training across thousands of encounters with diverse feedback and progressively greater autonomy. The paper, identified as arXiv:2608.07418v1, was announced as a new type on the arXiv preprint server. The method is presented as a way to enhance clinical reasoning, which relies on patient encounters involving history-taking, diagnostic hypothesis refinement, and management decisions under uncertainty.