ResidencyRL

ResidencyRL is a training method that puts AI through a simulated version of the training real medical residents go through. Instead of real patients, tens of thousands of fake patients are created with varied ages, genders, and personalities, and the AI repeatedly practices diagnosing, prescribing, and talking with them.

The key is how the training is scored. Rather than having the AI memorize a single correct answer, the training grades multiple aspects—whether the diagnosis was right, whether necessary questions were missed, whether the explanation was easy to understand—and lets the AI gradually adjust its own behavior toward better-scoring responses. It's less like memorizing an answer key and more like learning the ropes through repeated consultations and feedback on what went well and what didn't.

Google says it used this method to train its language model Gemini 3.5 Flash, improving its diagnostic, prescribing, and counseling abilities. However, even models trained this way reportedly haven't fully solved problems like fabricating false information or suggesting risky prescriptions.

The article reports that "Google used a reinforcement learning technique called ResidencyRL to train Gemini 3.5 Flash, improving its clinical consultation abilities." The name might suggest an actual hospital residency training program, but it's simply the name of a simulation-based training technique using virtual patients, unrelated to any real medical training system.