GATE: Graph-based Adaptive Tool Evolution Across Diverse Tasks

2025-02-20

AI Summary

Existing LLM-based tool construction frameworks suffer from low efficiency, poor cross-task generalization, and limited dynamic evolvability. To address these challenges, we propose a graph-structured adaptive tool evolution framework that unifies heterogeneous tasks—including open-domain reasoning, agent execution, and code generation—via hierarchical tool abstraction, online synthesis and pruning, and multi-task meta-evaluation-driven adaptive evolution. Our key contribution is the first-ever dynamic tool graph evolution mechanism that jointly optimizes across multiple tasks, automatically balancing tool scale, complexity, and functionality. Experimental results demonstrate significant improvements: 4.3× faster milestone completion in Minecraft; 9.23% and 10.03% performance gains on code generation and agent tasks, respectively—both substantially outperforming state-of-the-art baselines.

Technology Category

• Natural Language Processing: Code Generation / Program Synthesis from Natural Language
• Machine Learning: Evolutionary Learning
• Search and Optimization: Evolutionary Computation

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information

Abstract

Large Language Models (LLMs) have shown great promise in tool-making, yet existing frameworks often struggle to efficiently construct reliable toolsets and are limited to single-task settings. To address these challenges, we propose GATE (Graph-based Adaptive Tool Evolution), an adaptive framework that dynamically constructs and evolves a hierarchical graph of reusable tools across multiple scenarios. We evaluate GATE on open-ended tasks (Minecraft), agent-based tasks (TextCraft, DABench), and code generation tasks (MATH, Date, TabMWP). Our results show that GATE achieves up to 4.3x faster milestone completion in Minecraft compared to the previous SOTA, and provides an average improvement of 9.23% over existing tool-making methods in code generation tasks and 10.03% in agent tasks. GATE demonstrates the power of adaptive evolution, balancing tool quantity, complexity, and functionality while maintaining high efficiency. Code and data are available at url{https://github.com/ayanami2003/GATE}.

Problem

• Enhance tool-making efficiency across diverse tasks
• Construct reusable tools dynamically in multiple scenarios
• Improve performance in code and agent-based tasks

Innovation

• Graph-based adaptive tool evolution
• Hierarchical reusable tools construction
• Efficiency in multi-task tool-making