SOD: Step-wise On-policy Distillation for Small Language Model Agents

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Zhong, Qiyong, Zheng, Mao, Song, Mingyang, Lin, Xin, Sun, Jie, Jiang, Houcheng, Wang, Xiang, Fang, Junfeng
Natura: Preprint
Pubblicazione: 2026
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866911662541373440
author Zhong, Qiyong
Zheng, Mao
Song, Mingyang
Lin, Xin
Sun, Jie
Jiang, Houcheng
Wang, Xiang
Fang, Junfeng
author_facet Zhong, Qiyong
Zheng, Mao
Song, Mingyang
Lin, Xin
Sun, Jie
Jiang, Houcheng
Wang, Xiang
Fang, Junfeng
contents Tool-integrated reasoning (TIR) is difficult to scale to small language models due to instability in long-horizon tool interactions and limited model capacity. While reinforcement learning methods like group relative policy optimization provide only sparse outcome-level rewards. Recently, on-policy distillation (OPD) has gained popularity by supplying dense token-level supervision from a teacher on student-generated trajectories. However, our experiments indicate that applying OPD to TIR leads to a critical failure mode: erroneous tool calls tend to cascade across subsequent reasoning steps, progressively amplifying student-teacher divergence and rendering the teacher's token-level supervision increasingly unreliable. To address this, we propose SOD, a step-wise on-policy distillation framework for small language model agents, which adaptively reweights distillation strength at each step based on step-level divergence. Therefore, SOD can attenuate potentially misleading teacher signals in high-divergence regions while preserving dense guidance in well-aligned states. Experiments on challenging math, science, and code benchmarks show that SOD achieves up to 20.86% improvement over the second-best baseline. Notably, our 0.6B student achieves 26.13% on AIME 2025, demonstrating effective transfer of agentic reasoning to lightweight models. Our code is available at https://github.com/YoungZ365/SOD.
format Preprint
id arxiv_https___arxiv_org_abs_2605_07725
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle SOD: Step-wise On-policy Distillation for Small Language Model Agents
Zhong, Qiyong
Zheng, Mao
Song, Mingyang
Lin, Xin
Sun, Jie
Jiang, Houcheng
Wang, Xiang
Fang, Junfeng
Computation and Language
Artificial Intelligence
Tool-integrated reasoning (TIR) is difficult to scale to small language models due to instability in long-horizon tool interactions and limited model capacity. While reinforcement learning methods like group relative policy optimization provide only sparse outcome-level rewards. Recently, on-policy distillation (OPD) has gained popularity by supplying dense token-level supervision from a teacher on student-generated trajectories. However, our experiments indicate that applying OPD to TIR leads to a critical failure mode: erroneous tool calls tend to cascade across subsequent reasoning steps, progressively amplifying student-teacher divergence and rendering the teacher's token-level supervision increasingly unreliable. To address this, we propose SOD, a step-wise on-policy distillation framework for small language model agents, which adaptively reweights distillation strength at each step based on step-level divergence. Therefore, SOD can attenuate potentially misleading teacher signals in high-divergence regions while preserving dense guidance in well-aligned states. Experiments on challenging math, science, and code benchmarks show that SOD achieves up to 20.86% improvement over the second-best baseline. Notably, our 0.6B student achieves 26.13% on AIME 2025, demonstrating effective transfer of agentic reasoning to lightweight models. Our code is available at https://github.com/YoungZ365/SOD.
title SOD: Step-wise On-policy Distillation for Small Language Model Agents
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2605.07725