A new training method fixes a flaw that let AI agents fake confidence they had not earned.
Researchers studied on-policy self-distillation, a popular way to fine-tune AI agents after initial training. It works by showing a student model a "teacher" version of itself that gets extra privileged information, then grading the student against that teacher's output. The paper finds this backfires on multi-turn agents: the student learns to act as if it has information it never actually saw, and its real performance lags behind plain reinforcement learning, sometimes falling below an untrained base model. Their fix, called Privileged Self-Practice, keeps the privileged information but moves it out of the grading step and into the practice step. When a student mostly fails a task, an analyzer model writes a short hint, the student retries the task with that hint visible in the prompt, and only that retry gets used for training.
The distinction matters because it is easy to assume more information during training always helps. This paper shows that where that information sits, in the loss versus in the prompt, changes what the model actually learns, not just how well it scores. That is a useful caution for anyone building agent training pipelines around teacher-conditioned supervision.
Across AppWorld and SWE-bench Verified, tested on three different student models, the new method beat standard training every time, lifting task completion by up to 65% and issue resolution by up to 61%, though these are the paper's own benchmark numbers and have not yet been independently replicated.