AIAny
Icon for item

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

Evaluates a weak teacher's RL-induced policy shift on the student's own rollouts and amplifies verifier-supported updates so stronger models can learn from weaker supervisors and surpass them. It rescales only verifier-supported policy-gradient components to preserve optimization fixed points while accelerating learning, reducing student updates versus standard RL or distillation.

Introduction

Why this matters

Frontier-scale post-training (e.g., verifier-driven RL) is expensive to repeat on every new large model, and conventional distillation can lock a student into a weaker teacher's capacity ceiling. The central insight of this paper is that the useful supervision from weaker teachers can be extracted as a teacher policy shift relative to a reference checkpoint, evaluated on the student's own on-policy states, and then used to amplify only those verifier-supported gradient directions that help the student — accelerating progress without forcing the student to imitate the weak teacher.

Key Findings
  • Extracting the teacher's policy shift (difference in logit/policy space between post-RL and pre-RL checkpoints) and evaluating it on student rollouts yields a dense supervision signal that can be transferred across scales. This converts weak-model RL outcomes into a reusable signal for stronger students.
  • On-Policy Reverse Distillation (OPRD) rescales verifier-driven policy-gradient updates only along components supported by the teacher's shift; this preserves the stationary points of the underlying policy objective while boosting learning speed beyond the teacher's final policy.
  • Empirically, OPRD attains higher performance with fewer student updates than prior RL and distillation baselines in successive-model transfer and multi-teacher distillation, and it continues to improve students even when supervision sources are weaker overall.
  • Response-style analysis indicates OPRD-trained students stay closer in behavior to models trained with verifier-based RL alone than to their weak teachers, suggesting teacher guidance primarily accelerates the student's own optimization rather than redirecting it.
Who it's for & trade-offs

Great fit if you want to reuse cheap RL runs on smaller models as supervision for larger models (e.g., accelerate post-training of a new generation without rerunning expensive RL), or if you need a practical way to compose multiple weak supervision sources. OPRD reduces compute by avoiding sparse-reward RL on the target and by needing fewer student updates.

Look elsewhere if your verifier signal is unreliable or if teacher and student have fundamentally incompatible "thinking patterns"; OPRD relies on verifier-supported signals and paired teacher/reference checkpoints, and it can amplify spurious directions if the verifier does not correctly identify desirable behaviors. It also assumes access to both pre- and post-RL teacher checkpoints (or equivalent contrast pairs) to form a meaningful policy shift.

Where it fits (brief)

OPRD sits between imitation-style distillation and full sparse-reward RL: instead of copying a weaker teacher's final policy, it reuses the weak teacher's RL-induced policy shift as an implicit reward-like signal evaluated on the student's own states, combining verifier-based objectives with teacher guidance to speed up convergence without imposing the weak model's capacity ceiling.

Information

  • Websitearxiv.org
  • OrganizationsAffiliation: KAIST AI, Affiliation: Microsoft, Affiliation: University of Toronto, Affiliation: Mila, Affiliation: Université de Montréal, Affiliation: CIFAR AI Chair
  • AuthorsYoungrok Park, Sangmin Bae, Hojung Jung, Jongwoo Ko, Yunseon Choi, Young Jin Kim, Pashmina Cameron, Aaron Courville, Se-Young Yun
  • Published date2026/09/08

More Items

Trains a world-action model that predicts future visual states to guide zero-shot robotic manipulation; introduces CoAE, SVP, IDM and KASO to pretrain generative and action components from scratch on manipulation data, scaling up to 30,000 hours and improving zero-shot success.

Trains end-to-end driving without human trajectory supervision by decoupling perception and action: DriveVFM distills multiple frozen vision foundation models into a single camera backbone, and DriveRL trains a privileged closed-loop RL teacher whose rollouts supervise a camera-only planner, yielding state-of-the-art closed-loop benchmark results.

Explores a practical mechanism for recursive self-improvement by post-training LLMs: uses a routing harness to record agent executions and convert traces into curriculum-guided supervised fine-tuning and on-policy distillation data, closing an evaluation-selection-update loop and improving benchmark performance.