Researchers refine on-policy distillation for smaller models

Several arXiv papers propose variants of on-policy distillation, a training approach that supervises student models on trajectories they generate themselves. The methods include Fisher-projected targets for vision-language models, future-trajectory validation for multi-turn agents, sparse probing with outcome-calibrated targets, and recoverability-aware handling of erroneous prefixes. Reported evaluations include gains for FutureBridge-OPD on ALFWorld, WebShop and ScienceWorld, and recoverability-aware control on AIME and GPQA-Diamond benchmarks.