Crimson AI NewsA CrimsonLingua Network service
EN ع
← Back to news
Research paper Hugging Face

Study: On-Policy Distillation Works by Suppressing Low-Probability Tokens, Not Teacher Guidance

AI By Crimson AI Hugging Face Papers 1 September 2026 · 00:00 2 views
Share: X Telegram

A new Hugging Face paper reveals that on-policy distillation (OPD) improves reasoning mainly by suppressing low-probability tokens, not by teacher guidance. The authors propose OPSA, a teacher-free method that uses entropy-adaptive negative advantages, achieving substantial gains on reasoning benchmarks.

Study: On-Policy Distillation Works by Suppressing Low-Probability Tokens, Not Teacher Guidance

Key points

On-policy distillation (OPD) has been proposed as an alternative to reinforcement learning with verifiable rewards (RLVR), offering dense token-level supervision. However, a new paper from Hugging Face questions the source of OPD's effectiveness, suggesting that its gains may not come from teacher guidance at all.

The researchers quantitatively analyzed teacher supervision during OPD training and found substantial noise, which increases with teacher scale. Surprisingly, the student policy is insensitive to this noise, converging to comparable performance whether noisy supervision is retained or removed. This raises the question: does OPD actually distill?

By analyzing what drives OPD's gains, the authors found that learning concentrates on low log-probability tokens. Using a single fixed negative advantage matches the performance of teacher-provided ones, suggesting that OPD works largely by suppressing low-probability tokens, which requires no teacher.

Based on these findings, the paper introduces On-Policy Self-Adaptation (OPSA), a supervision-free method that uses entropy-adaptive negative advantages. OPSA assigns stronger learning signals to high-entropy positions, suppressing tail tokens and evenly redistributing probability mass among head tokens.

Compared with the base Qwen3-1.7B, OPSA improves Avg@32 by 35.41 points on AIME24 (a 263% relative gain) and more than doubles Pass@32 across all three benchmarks. It also outperforms OPD by 16.77 points in Avg@32 on AIME24. Extensive experiments across model families and tasks demonstrate its effectiveness and generalizability.

MethodAvg@32 on AIME24Relative Gain vs Base
Base Qwen3-1.7BBaseline-
OPSA+35.41+263%
OPSA vs OPD+16.77-
Source
Hugging Face · Hugging Face Papers
Related news
Research paper
Hugging Face 1 Sep 2026

Hugging Face Unveils CogEvol: Single-Pass Learning Environment Generation

CogEvol, a new family of models from Hugging Face, generates structured learning artifacts like slides and interactive HTML pages...

1
Research paper
Hugging Face 1 Sep 2026

PaperGym: Turning Papers into Training Environments for AI Research Planning

Hugging Face researchers introduce PaperGym, a framework that converts scientific papers into reinforcement learning environments...

1
Research paper
Hugging Face 1 Sep 2026

Hugging Face Unveils Lucida: A New Pipeline for Composable Real-to-Sim Scene Modeling

Lucida, a new method from Hugging Face, improves composable indoor scene reconstruction by redistributing pipeline requirements ac...

0