AIAny
Icon for item

Does On-Policy Distillation Really Distill? From Noisy Teacher to Self-Improvement

Analyzes on-policy distillation for LLM fine-tuning, shows teacher token-level supervision is often noisy and not the main driver of gains, and introduces OPSA, a supervision-free, entropy-adaptive method that suppresses low-probability tokens to improve downstream accuracy.

Introduction

Why this matters: dense token-level supervision from a strong teacher has been proposed as an efficient alternative to sparse outcome rewards for improving LLMs, but student trajectories are often off-policy for the teacher. The paper's core insight is counterintuitive: teacher-provided advantages contain substantial, scale-dependent noise, yet students still improve because learning concentrates on suppressing low-probability tokens — a behavior that needs no teacher.

Key Findings
  • High teacher noise: a large fraction of token-level advantages are inconsistent with answer correctness, and the noise rate increases with teacher scale; the biggest teachers often assign overwhelmingly negative advantages regardless of correctness, indicating distributional mismatch.
  • Which tokens matter: improvements come mainly from the small subset of low-log-probability (tail) tokens sampled by the student; high-probability tokens contribute little to gradients or learning.
  • Which signals matter: negative advantages dominate; replacing teacher advantages with a single fixed negative value reproduces much of OPD's gains, showing suppression of tail tokens—not teacher knowledge transfer—is the primary mechanism.
  • Method: proposes On-Policy Self-Adaptation (OPSA), a supervision-free scheme that assigns entropy-adaptive negative advantages to low-probability tokens, sharpening low-entropy positions while preserving exploration at high-entropy positions.
Who it's for and tradeoffs

Great fit if you study LLM fine-tuning dynamics, token-level RL methods, or want lightweight alternatives to teacher-based distillation: the paper gives diagnostic analyses, ablations, and a practical supervision-free method that substantially improves benchmarks like AIME24 on mid-sized models. Look elsewhere if your setting requires strict behavior cloning or you rely on white-box teacher logits and shared vocabularies: OPSA assumes you want to reshape student distributions without external teacher access and may not replace approaches that require explicit teacher alignment for specific behavior replication.

Where it fits

This work reframes on-policy distillation as largely a distribution-shaping operation rather than faithful knowledge transfer, and positions OPSA as a cheaper, generalizable replacement for OPD in many LLM post-training pipelines that aim to reduce tail-token errors and increase pass@k metrics.

Information

  • Websitearxiv.org
  • OrganizationsAffiliation: Department of Computer Science, Purdue University, USA
  • AuthorsYi Ding, Ruqi Zhang
  • Published date2026/08/31

More Items

Turns each research paper into a training environment to generate verifiable research plans by synthesizing questions from goals/background and deriving evaluation criteria from methods/experiments. Key features: four-stage extraction that reduces criterion leakage to 3.7%, a two-stage rubric-centered training (self-distillation then GRPO), and the PaperGym-20k corpus with two held-out benchmarks.

Trains LLM agents to proactively edit and manage their working context for long-horizon tasks using an expanded toolset (planning, long-term memory, soft offloading) and a fine-grained RL algorithm that identifies critical edits and assigns action-level credit. Improves accuracy while keeping contexts compact on long-context QA and deep search.

Proposes VLAct, a representation-centric continued pre-training method for Vision-Language-Action models that preserves VLM priors and enforces cross-embodiment action semantics to turn limited robot trajectories into transferable visual-action representations; shows strong gains and sample efficiency on multiple VLA benchmarks using modest compute.