ResidencyRL Trains Clinical Agents Through Simulated Consultations of Up to 60 Turns, Raising Diagnostic Accuracy on Adversarial Cases to 88%

Static medical question answering primarily measures knowledge, but cannot verify whether a model knows when to ask follow-up questions, order tests, revise a differential diagnosis, or end a consultation. ResidencyRL reframes clinical reasoning as a long-horizon sequential decision-making problem. Using Gemini 3.5 Flash as the policy model, each training trajectory includes up to 60 dialogue turns and eight structured tool calls. While conversing with a simulated patient, the agent must also submit a diagnosis, management plan, and clinical documentation.

The training environment begins with a generator that creates cases spanning five clinical domains, 81 chief complaints, and varying levels of complexity. It then introduces adversarial behaviors, such as withholding information, resisting recommendations, or pressuring the agent to provide inappropriate treatment. Gemini 3.1 Pro is used for case generation and automated evaluation. Rewards jointly assess diagnosis, management, history-taking completeness, communication, documentation quality, and style, with additional penalties for fabricating medical history, recommending contraindicated treatments, underestimating urgency, or missing red-flag symptoms. The policy is ultimately trained with GRPO by comparing the relative rewards of multiple rollouts from the same scenario.

Across 200 held-out adversarial scenarios, the paper reports that diagnostic accuracy rose from 81% for the base model to 88%, while the rate of missed red-flag symptoms fell by 31%. The gains also transferred to 120 three-visit cases in AMIE Mx, oncology specialty cases, and external environments including AgentClinic and CRAFT-MD. Most external benchmarks showed improvements in the same direction, although not every metric was accompanied by a finding of statistical significance. In a separate set of 97 valid cases evaluated with the AMIE telehealth harness, blinded clinicians preferred the ResidencyRL-trained version in 87.6% of pairwise comparisons.

The significance of this work lies in shifting the optimization target for medical models from a “single answer” to the full consultation workflow, but the evidence remains limited to simulated environments. Both the patients and evaluators are LLM-driven and may share blind spots, while some evaluations also use models from the same family as judges. Google says that specific implementation details depend on internal infrastructure and that the code will not be released because of medical safety considerations, so the training process cannot currently be reproduced independently. The next priorities for engineering and research teams should be prospective testing in real-world workflows, cross-model evaluator calibration, and determining whether simulated rewards induce new forms of clinical specification gaming.