arxiv:2609.08798

Eliciting Weak-to-Strong Generalization with On-Policy Reverse Distillation

Published on Sep 8

· Submitted by

Sangmin Bae on Sep 9

KAIST AI

Upvote

Authors:

Jongwoo Ko ,

Abstract

On-Policy Reverse Distillation enables stronger models to exceed weak supervisors by amplifying verifier-supported policy gradients along the teacher's shift direction, accelerating optimization without imposing capacity limits.

Generated by thinkingmachines/Inkling-Small

Weak-to-strong generalization asks whether stronger models can learn from weaker supervisors and surpass them. This question is particularly important for successive model generations and multi-domain consolidation, where repeating frontier-scale post-training from scratch can be prohibitively expensive. Yet conventional distillation treats the weak teacher as an optimization target, potentially imposing its capacity ceiling on the student. We introduce On-Policy Reverse Distillation (OPRD), which evaluates the teacher's policy shift relative to its reference policy on student rollouts and amplifies the component of the student's verifier-driven policy gradient along that direction. By rescaling only verifier-supported updates, OPRD preserves the stationary points of policy optimization while accelerating learning beyond the teacher. In both successive model transfer and multi-teacher distillation, OPRD achieves higher performance with fewer student updates than existing RL and distillation approaches. Response-style analysis shows that OPRD students remain closer to models trained with verifier-based RL alone than to their weak teachers, suggesting that teacher guidance accelerates rather than redirects the student's own optimization. Results in conventional strong-to-weak distillation further demonstrate that OPRD effectively combines verifier-driven policy optimization with teacher guidance regardless of capacity ordering.

View arXiv page View PDF GitHub 4 Add to collection

Community

bsmn0223

Paper submitter about 13 hours ago

Contributions

Our key contributions are as follows:

Weak-to-Strong Generalization. We study how post-training gains from weaker models can be transferred to stronger students in two practical scenarios: successive model transfer and multi-domain consolidation. We identify the central challenge as exploiting these gains without making either the weak policy or its policy shift a separate optimization target.
On-Policy Reverse Distillation. We introduce OPRD, which evaluates a weak teacher's policy shift relative to its reference policy on student rollouts and amplifies the component of the student's verifier-driven policy gradient along that direction. Because OPRD only rescales verifier-supported updates, it accelerates the student's own optimization while preserving its stationary points, allowing the student to move beyond the teacher.
Empirical Evaluation and Analysis. Across successive-model and multi-teacher settings, OPRD reaches the final performance of competing methods substantially earlier and ultimately outperforms both RL and distillation baselines. We further confirm that these gains extend to conventional strong-to-weak distillation. We also compare against recent weak-to-strong methods, analyze the design and dynamics of teacher guidance, examine practical challenges and mitigations, and study student reasoning and response style under teacher guidance.