Researchers refine on-policy distillation for AI agents

Three arXiv papers propose new methods for on-policy distillation, a training approach that supervises student models on trajectories they actually visit. FutureBridge-OPD tests whether a short teacher intervention at high-disagreement states improves later student trajectories, reporting average gains of 16.6 points over vanilla OPD and 7.6 points over TCOD on ALFWorld, WebShop and ScienceWorld in a Qwen3-32B-to-Qwen3-1.7B setup. SPOT uses sparse probing and verifier-scored continuations to decide where and what to distill, while a counterfactual recoverability method replays error states to decide whether to continue, roll back or supervise conventionally.