New papers target weak points in on-policy distillation

A set of arXiv papers proposes new methods for on-policy distillation, a training approach that supervises student models on trajectories they generate themselves. The papers address several failure modes: sparse or poorly calibrated teacher guidance, uncertainty about whether student errors are recoverable, degradation when privileged context is reintroduced, and whether teacher interventions help later agent trajectories. Reported experiments include gains on AIME, GPQA-Diamond, ALFWorld, WebShop and ScienceWorld, but the evidence is limited to the authors' evaluations.