AutoTool: Efficient Tool Selection for Large Language Model Agents

TL;DR

AutoTool addresses the high inference cost of tool selection in LLM agents by modeling tool usage as structured inertia with a dynamic Tool Inertia Graph (TIG). It performs inertial tool prediction via Comprehensive Inertia Potential Score (CIPS) and fills tool inputs through hierarchical parameter filling, reducing LLM calls by roughly $15$–$25\%$ and token usage by $10$–$40\%$ across diverse tasks, while preserving progress rates. The approach is validated on AlfWorld, ScienceWorld, and ToolQuery-Academic, and shows generalization across models and domains, including a macro-level ToolBench analysis confirming low-entropy, predictable tool sequences. AutoTool thus provides a training-free, graph-based acceleration that improves efficiency of inference-heavy LLM agent pipelines without sacrificing task performance, with mechanisms for fault tolerance and online adaptation. Future directions include dynamic inertia windows and tighter integration of semantic information to further enhance decision reliability in dynamic environments.

Abstract

Large Language Model (LLM) agents have emerged as powerful tools for automating complex tasks by leveraging the reasoning and decision-making abilities of LLMs. However, a major bottleneck in current agent frameworks lies in the high inference cost of tool selection, especially in approaches like ReAct that repeatedly invoke the LLM to determine which tool to use at each step. In this work, we propose AutoTool, a novel graph-based framework that bypasses repeated LLM inference by exploiting a key empirical observation: tool usage inertia - the tendency of tool invocations to follow predictable sequential patterns. AutoTool constructs a directed graph from historical agent trajectories, where nodes represent tools and edges capture transition probabilities, effectively modeling the inertia in tool selection. It further integrates parameter-level information to refine tool input generation. By traversing this structured representation, AutoTool efficiently selects tools and their parameters with minimal reliance on LLM inference. Extensive experiments across diverse agent tasks demonstrate that AutoTool reduces inference costs by up to 30% while maintaining competitive task completion rates, offering a practical and scalable enhancement for inference-heavy frameworks. Our work highlights the promise of integrating statistical structure into LLM agent design for greater efficiency without sacrificing performance.

Figures (11)

• Figure 1: The distribution of successor tools for different tools or tool sequences.
• Figure 2: An overview of the AutoTool framework. This figure illustrates our proposed workflow, which initiates with the Inertia Sensing module predicting the next likely tool by exploiting historical usage patterns. If a high-confidence tool is identified, the Parameter Filling module then populates its required parameters. Only when both stages succeed is the tool executed directly via the inertial path, bypassing a costly LLM call. Otherwise, the system reverts to a standard LLM invocation.
• Figure 3: ToolBench Inertia Analysis. We calculated the frequency distribution of successor tools for each tool and derived the proportion of its top-$k$ successor tools among all its successor tools.
• Figure 4: Case study for scienceworld
• Figure 5: Failure Scenario 1