Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Feng, Zihao, Wang, Xiaoxue, Wu, Bowen, Cao, Hailong, Zhao, Tiejun, Yu, Qun, Wang, Baoxun
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2509.14718
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866912591550349312
author Feng, Zihao
Wang, Xiaoxue
Wu, Bowen
Cao, Hailong
Zhao, Tiejun
Yu, Qun
Wang, Baoxun
author_facet Feng, Zihao
Wang, Xiaoxue
Wu, Bowen
Cao, Hailong
Zhao, Tiejun
Yu, Qun
Wang, Baoxun
contents While reinforcement learning (RL) is increasingly used for LLM-based tool learning, its efficiency is often hampered by an overabundance of simple samples that provide diminishing learning value as training progresses. Existing dynamic sampling techniques are ill-suited for the multi-task structure and fine-grained reward mechanisms inherent to tool learning. This paper introduces Dynamic Sampling with Curriculum Learning (DSCL), a framework specifically designed to address this challenge by targeting the unique characteristics of tool learning: its multiple interdependent sub-tasks and multi-valued reward functions. DSCL features two core components: Reward-Based Dynamic Sampling, which uses multi-dimensional reward statistics (mean and variance) to prioritize valuable data, and Task-Based Dynamic Curriculum Learning, which adaptively focuses training on less-mastered sub-tasks. Through extensive experiments, we demonstrate that DSCL significantly improves training efficiency and model performance over strong baselines, achieving a 3.29\% improvement on the BFCLv3 benchmark. Our method provides a tailored solution that effectively leverages the complex reward signals and sub-task dynamics within tool learning to achieve superior results.
format Preprint
id arxiv_https___arxiv_org_abs_2509_14718
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle ToolSample: Dual Dynamic Sampling Methods with Curriculum Learning for RL-based Tool Learning
Feng, Zihao
Wang, Xiaoxue
Wu, Bowen
Cao, Hailong
Zhao, Tiejun
Yu, Qun
Wang, Baoxun
Machine Learning
Computation and Language
While reinforcement learning (RL) is increasingly used for LLM-based tool learning, its efficiency is often hampered by an overabundance of simple samples that provide diminishing learning value as training progresses. Existing dynamic sampling techniques are ill-suited for the multi-task structure and fine-grained reward mechanisms inherent to tool learning. This paper introduces Dynamic Sampling with Curriculum Learning (DSCL), a framework specifically designed to address this challenge by targeting the unique characteristics of tool learning: its multiple interdependent sub-tasks and multi-valued reward functions. DSCL features two core components: Reward-Based Dynamic Sampling, which uses multi-dimensional reward statistics (mean and variance) to prioritize valuable data, and Task-Based Dynamic Curriculum Learning, which adaptively focuses training on less-mastered sub-tasks. Through extensive experiments, we demonstrate that DSCL significantly improves training efficiency and model performance over strong baselines, achieving a 3.29\% improvement on the BFCLv3 benchmark. Our method provides a tailored solution that effectively leverages the complex reward signals and sub-task dynamics within tool learning to achieve superior results.
title ToolSample: Dual Dynamic Sampling Methods with Curriculum Learning for RL-based Tool Learning
topic Machine Learning
Computation and Language
url https://arxiv.org/abs/2509.14718