Prompt Breadth and Rollout Refresh Interact in On-Policy Distillation

Abstract:How many prompts does on-policy distillation (OPD) need, and how does the answer depend on the student policies that generate its training responses? We study these two controls jointly: prompt breadth and rollout refresh. A 3x3 mathematical-reasoning experiment fixes 14,080 trajectories and 110 optimizer updates while varying the prompt bank and the number of response-generating policy snapshots. With ten snapshots, eight prompts reach 24.09% average accuracy, close to 24.51% for 14,080 distinct prompts. With responses frozen at the initial policy, however, increasing breadth lowers accuracy from 21.16% to 19.05%; under per-update refresh, it raises accuracy from 23.61% to 25.57%. The resulting interaction is 4.07 percentage points, with a 95% question-paired interval of [2.00, 6.28]. Matched comparisons under two teachers reveal a second reversal: the periodic models have higher short-budget accuracy and answer completion, but frozen-response models overtake in average accuracy at a 32K output limit, using 1.7-1.8x as many response tokens. These results show that prompt efficiency in OPD can depend on both refresh and inference budget.

Computation and Language (cs.CL); Artificial Intelligence (cs.AI)

arXiv:2609.25048 [cs.CL]

(or arXiv:2609.25048v1 [cs.CL] for this version)

https://doi.org/10.48550/arXiv.2609.25048 Focus to learn more arXiv-issued DOI via DataCite

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.