AI learns clinical judgment from simulated patients

🤖 AI learns clinical judgment from simulated patients

Researchers introduced ResidencyRL, an RL training method for clinical AI agents based on Gemini 3.5 Flash. The agent undergoes a "residency" through simulated appointments: up to 60 dialogue turns, 8 tool calls per trajectory. Training on 49,870 telemedicine cases, 81 diagnoses.

🌍 Diagnostic accuracy on complex tests +7 p.p. (88% vs 81%). Missed critical questions −22 p.p. Blind expert doctors chose the trained agent in 87.6% of comparisons.

👤 Currently an auxiliary tool: history taking, checking against guidelines. Far from autonomous practice — no physical examinations or laboratory tests.

Source: https://arxiv.org/abs/2608.07418