Skill-Conditioned Gated Self-Distillation for LLM Reasoning

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Huang, Jiazhen, Chen, Xiao, Luo, Xiao, Dai, Yong, Hu, Senkang, Zhao, Yuzhi
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866916056390434816
author Huang, Jiazhen
Chen, Xiao
Luo, Xiao
Dai, Yong
Hu, Senkang
Zhao, Yuzhi
author_facet Huang, Jiazhen
Chen, Xiao
Luo, Xiao
Dai, Yong
Hu, Senkang
Zhao, Yuzhi
contents On-policy self-distillation (SD) improves LLM reasoning by using teacher-side privileged information (PI) to turn sparse verifier outcomes into dense token-level supervision. Existing methods usually assume trusted PI, such as reference answers or successful traces. We ask whether PI can instead come from an experience-derived skill bank, where retrieved skills are compact and reusable but may also be irrelevant or misleading. We propose Skill-Conditioned Gated Self-Distillation (SGSD), which formulates skill-based SD as teacher hypothesis validation rather than unconditional imitation. SGSD retrieves skill-mistake pairs, constructs a multi-teacher pool, and lets all skill-conditioned teachers score the same plain-prompt student rollout. The verifier validates each teacher's polarity: supporting a success or suppressing a failure gives positive supervision, while the opposite stance is reversed. A robust gated objective then distills informative teacher-student disagreements while suppressing uncertain or extreme signals. Experiments on multiple mathematical reasoning benchmarks show that SGSD consistently improves over GRPO and remains competitive with answer-conditioned OPSD under a weaker PI assumption. For example, on Qwen3-1.7B, SGSD outperforms GRPO by 6.2% and OPSD by 1.7% on average on AIME24, AIME25, and HMMT25. Our code is available at https://github.com/walawalagoose/SGSD.
format Preprint
id arxiv_https___arxiv_org_abs_2605_28791
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Skill-Conditioned Gated Self-Distillation for LLM Reasoning
Huang, Jiazhen
Chen, Xiao
Luo, Xiao
Dai, Yong
Hu, Senkang
Zhao, Yuzhi
Computation and Language
Artificial Intelligence
On-policy self-distillation (SD) improves LLM reasoning by using teacher-side privileged information (PI) to turn sparse verifier outcomes into dense token-level supervision. Existing methods usually assume trusted PI, such as reference answers or successful traces. We ask whether PI can instead come from an experience-derived skill bank, where retrieved skills are compact and reusable but may also be irrelevant or misleading. We propose Skill-Conditioned Gated Self-Distillation (SGSD), which formulates skill-based SD as teacher hypothesis validation rather than unconditional imitation. SGSD retrieves skill-mistake pairs, constructs a multi-teacher pool, and lets all skill-conditioned teachers score the same plain-prompt student rollout. The verifier validates each teacher's polarity: supporting a success or suppressing a failure gives positive supervision, while the opposite stance is reversed. A robust gated objective then distills informative teacher-student disagreements while suppressing uncertain or extreme signals. Experiments on multiple mathematical reasoning benchmarks show that SGSD consistently improves over GRPO and remains competitive with answer-conditioned OPSD under a weaker PI assumption. For example, on Qwen3-1.7B, SGSD outperforms GRPO by 6.2% and OPSD by 1.7% on average on AIME24, AIME25, and HMMT25. Our code is available at https://github.com/walawalagoose/SGSD.
title Skill-Conditioned Gated Self-Distillation for LLM Reasoning
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2605.28791