One training example recovers 71.5% of what full-data on-policy distillation reaches, and 16 match it outright.

arXiv 2609.04172 trains OPD with a single query and finds it keeps improving for hundreds of steps across task domains and model families. Measuring state coverage, the fraction of full-data states a query set's rollouts reach, one query hits 71.5% and 16 semantically distinct queries reach 98.9%. Alignment slows at the same rate either way. The authors summarize it as OPD being data-overfed and algorithm-starved, and content-light templates plus off-domain WildChat queries also approached the real-query baseline (arXiv).