概念
distillation on-policy self-distillation training
创建2026-04-15
更新2026-04-15
阅读量级4 分钟
概念导读

On-Policy Self-Distillation for Reasoning Compression (OPSDC) is a training method that compresses LLM reasoning traces by conditioning the same model on a "be concise" instruction to generate teacher logits, then training the student (unconditioned) to minimize per-token reverse KL divergence on its own rollouts. No ground-truth answers are needed.

  1. Overview
  2. Core Mechanism
  3. Key Results
  4. Relationships
  5. See Also

On-Policy Self-Distillation for Reasoning Compression (OPSDC)

On-Policy Self-Distillation for Reasoning Compression (OPSDC) is a training method that compresses LLM reasoning traces by conditioning the same model on a "be concise" instruction to generate teacher logits, then training the student (unconditioned) to minimize per-token reverse KL divergence on its own rollouts. No ground-truth answers are needed.

Overview

LLM reasoning models (e.g., chain-of-thought) often produce verbose traces with unnecessary deliberation. OPSDC compresses these traces by teaching the model to produce shorter, equally accurate reasoning. The key insight: the same model can serve as its own teacher by conditioning it differently — with a conciseness instruction for teacher logits, without it for student rollouts.

Core Mechanism

  1. Dual Conditioning: The same model is used twice:
  2. Teacher mode: Conditioned on "be concise" instruction prepended to the prompt. Produces logits pi_teacher(y|x, "be concise").
  3. Student mode: Standard prompting without the conciseness instruction. Generates rollouts pi_student(y|x).

  4. On-Policy Rollouts: The student generates rollouts from its own policy (on-policy). No offline teacher traces are needed.

  5. Per-Token Reverse KL Minimization: At each token position in the student's rollout, minimize KL(pi_student || pi_teacher). This encourages the student to match the teacher's more compact token distribution.

  6. No Ground Truth Needed: Since the objective is self-consistency (match the concise version of yourself), no external ground-truth answers or solutions are required. The model learns to self-compress.

Key Results

Model Token Reduction Accuracy Gain (MATH-500)
Qwen3-8B 57% +9–16 pts
Qwen3-14B 59% +9–16 pts

OPSDC achieves substantial compression (over half the tokens removed) while actually improving accuracy, suggesting that verbose reasoning traces contain noise that self-distillation filters out.

Relationships

  • Self-distillation variant of OPD: the teacher is the same model with different conditioning.
  • Related to OPSD but specifically targets reasoning compression rather than general capability transfer.
  • Complementary to SRPO: OPSDC shows that self-conditioning can serve as the "teacher" signal, while SRPO uses an external teacher for failed samples.
  • The per-token reverse KL objective connects to the theoretical framework in G-OPD.

See Also

[src: raw/papers/2026/03/2603.05433.md]