Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ToolACE-DEV: Self-Improving Tool Learning via Decomposition and EVolution

Xu Huang, Weiwen Liu, Xingshan Zeng, Yuefeng Huang, Xinlong Hao, Yuxian Wang, Yirong Zeng, Chuhan Wu, Yasheng Wang, Ruiming Tang, Defu Lian

TL;DR

ToolACE-DEV introduces a self-evolving framework for tool learning that decomposes tool usage into tool documentation adaption, query-aware tool generation, and tool invocation, then enables autonomous self-improvement through iterative data generation and model updating. By training on tool definitions and enabling generation of candidate tools, the approach reduces reliance on expensive advanced models and addresses data-compatibility and privacy concerns. Empirical results show that an 8B LLaMA-based model with self-evolution can achieve state-of-the-art-like tool-invocation accuracy on BFCL and exhibits strong data-efficiency, especially on live, user-contributed tool sets. Self-evolution yields progressive gains over rounds, with larger improvements on harder tasks and diminishing returns as confidence grows, highlighting both the potential and limits of autonomous tool-learning for lightweight LLMs.

Abstract

The tool-using capability of large language models (LLMs) enables them to access up-to-date external information and handle complex tasks. Current approaches to enhancing this capability primarily rely on distilling advanced models by data synthesis. However, this method incurs significant costs associated with advanced model usage and often results in data compatibility issues, led by the high discrepancy in the knowledge scope between the advanced model and the target model. To address these challenges, we propose ToolACE-DEV, a self-improving framework for tool learning. First, we decompose the tool-learning objective into sub-tasks that enhance basic tool-making and tool-using abilities. Then, we introduce a self-evolving paradigm that allows lightweight models to self-improve, reducing reliance on advanced LLMs. Extensive experiments validate the effectiveness of our approach across models of varying scales and architectures.

ToolACE-DEV: Self-Improving Tool Learning via Decomposition and EVolution

TL;DR

ToolACE-DEV introduces a self-evolving framework for tool learning that decomposes tool usage into tool documentation adaption, query-aware tool generation, and tool invocation, then enables autonomous self-improvement through iterative data generation and model updating. By training on tool definitions and enabling generation of candidate tools, the approach reduces reliance on expensive advanced models and addresses data-compatibility and privacy concerns. Empirical results show that an 8B LLaMA-based model with self-evolution can achieve state-of-the-art-like tool-invocation accuracy on BFCL and exhibits strong data-efficiency, especially on live, user-contributed tool sets. Self-evolution yields progressive gains over rounds, with larger improvements on harder tasks and diminishing returns as confidence grows, highlighting both the potential and limits of autonomous tool-learning for lightweight LLMs.

Abstract

The tool-using capability of large language models (LLMs) enables them to access up-to-date external information and handle complex tasks. Current approaches to enhancing this capability primarily rely on distilling advanced models by data synthesis. However, this method incurs significant costs associated with advanced model usage and often results in data compatibility issues, led by the high discrepancy in the knowledge scope between the advanced model and the target model. To address these challenges, we propose ToolACE-DEV, a self-improving framework for tool learning. First, we decompose the tool-learning objective into sub-tasks that enhance basic tool-making and tool-using abilities. Then, we introduce a self-evolving paradigm that allows lightweight models to self-improve, reducing reliance on advanced LLMs. Extensive experiments validate the effectiveness of our approach across models of varying scales and architectures.
Paper Structure (21 sections, 7 equations, 7 figures, 3 tables)

This paper contains 21 sections, 7 equations, 7 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Tool Learning
  4. Self Evolution of LLMs
  5. Methodology
  6. Task Definitions
  7. Tool Documentation Adaption
  8. Query-Aware Tool Generation and Invocation
  9. Self-Evolution
  10. Experiments
  11. Experimental Settings
  12. Main Results
  13. Performance of Self-Evolution
  14. Ablation Study
  15. Effectiveness on various models
  16. ...and 6 more sections

Figures (7)

  • Figure 1: Overall framework of the self-evolving paradigm. (a) Tool documentation adaption, aiming to enhance the understanding of tools; (b) Query-aware tool generation and invocation, equipping the model with self-evolving abilities; (c) Self-evolution, where the model first generates candidate tools and then generates tool invocation, forming the self-training data.
  • Figure 2: Relative Improvements of Self-Evolution. The backbone model is LLaMA-3.1-8B-Instruct.
  • Figure 3: Ablation study of different training objectives on various models.
  • Figure 4: Effects of self-evolution on various models.
  • Figure 5: Illustration of the tool invocation task. The blue part represents the expected output.
  • ...and 2 more figures