AIAny
Icon for item

Weak-to-Strong Generalization via Direct On-Policy Distillation

Transfers RL-induced policy shifts from a smaller 'weak' teacher to a stronger target by using the teacher's post-/pre-RL log-ratio as a dense implicit reward applied on the student's on-policy states. Enables reuse of RL supervision without running RL rollouts on the target, improving sample/time efficiency.

Introduction

Repetition of expensive reward-driven fine-tuning for each new, larger model becomes a major bottleneck as models scale. The core insight of Direct On-Policy Distillation (Direct-OPD) is to treat the policy change produced by running RL on a cheaper, smaller “weak” model as an implicit, dense reward signal that can be applied to a stronger target without re-running sparse-reward RL.

Key Findings
  • Directly computes the log-ratio between a weak model's post-RL and pre-RL policies and uses that scalar as a dense reward for the student on the student's own on-policy states — so what: it captures which actions RL made more or less likely, abstracting away the weak model's capacity limits.
  • Substantially faster and more compute-efficient than running RL on the target: example result shows Qwen3-1.7B improving from 48.3% to 58.3% on AIME 2024 in ≈4 hours on 8 A100 GPUs — so what: practical gains for large-model post-training without full RL rollouts.
  • Outperforms step-matched direct RL and supports sequential composition of multiple policy shifts — so what: you can chain multiple weak-to-strong transfers to accumulate improvements.
Who it's for and tradeoffs
  • Great fit if you need to transfer sparse-reward RL gains from cheap, smaller models to expensive large models, or you want to avoid running many target-model rollouts while preserving RL-derived behavior changes.
  • Look elsewhere if the weak teacher is poorly aligned with target tasks or the RL improvements rely heavily on exploration patterns that do not transfer via log-ratio signals; the method assumes the teacher's policy shift encodes transferable preferences rather than model-specific failure modes.
How it works (brief)

Direct-OPD requires a paired checkpoint (pre-RL, post-RL) of a weak teacher. For states sampled on the student's on-policy rollout, it evaluates the teacher's probability ratio log(pi_post / pi_pre) for the action taken and uses that scalar as an implicit reward to train the student. This reframes the weak teacher's RL outcome as dense supervision mapped onto the student's state distribution, avoiding sparse-reward optimization on the target while preserving the direction of the RL-induced policy shift.

Information

  • Websitearxiv.org
  • AuthorsShiyuan Feng, Huan-ang Gao, Haohan Chi, Hanlin Wu, Zhilong Zhang, Zheng Jiang, Bingxiang He, Wei-Ying Ma, Ya-Qin Zhang, Hao Zhou
  • Published date2026/07/06

More Items

Introduces KronQ, a post-training quantization framework that incorporates gradient covariance via a Kronecker‑factored Hessian to guide input/output weight rotations and sensitivity-driven mixed-precision allocation. Demonstrates stable 2-bit weight-only quantization on LLaMA-3-70B (7.93 PPL).

Stabilizes on-policy policy distillation by dynamically constructing a proximal teacher that controls gradient variance. Provides theoretical global convergence and monotonic improvement bounds, and shows improved training stability, sample efficiency, and final performance on mathematical reasoning tasks with zero extra compute overhead.

Explores unsupervised visual pretraining on visually rich documents to improve language-model intelligence; shows visual-pretrained models outperform text-only counterparts on the same corpora. Key aspects: direct use of images/layouts (no OCR-only pipeline), scalable across backbones and benchmarks.