AutoTool: Efficient Tool Selection for Large Language Model Agents

📝 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.

💡 Analysis

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.

📄 Content

Large Language Models (LLMs) have experienced explosive growth, demonstrating impressive capabilities in various tasks (Achiam et al. 2023;Dubey et al. 2024b;Yang et al. 2025) from natural language understanding (Devlin et al. 2019) and reasoning (Guo et al. 2025) to automation of complex workflows (Wang et al. 2025). LLM-powered agents, which leverage these capabilities for interactive (Yi et al. 2024) and decision-making (Wei et al. 2025) tasks, have become increasingly prevalent in numerous domains, including software development (Jin et al. 2024), intelligent personal assistants (Li et al. 2024) and scientific research automation (Team et al. 2025).

However, despite their versatility, a significant drawback of current LLM-based agents is the substantial computational overhead (Kim et al. 2025), particularly evident in frameworks like ReAct (Yao et al. 2023b) that involve many LLM inferences. Among these inference processes, one major goal is to repeatedly infer appropriate tools to use based on dynamic contexts, leading to high inference costs and latency. Given these challenges, a natural question arises: Can we utilize statistical methods instead of relying heavily on LLM inference to select tools efficiently?

Addressing this issue, we empirically observe a critical phenomenon termed tool usage inertia, where tool selections demonstrate predictable sequential patterns. For instance, when an agent searches in an academic database, the invocation of AuthorNodeCheck is often followed by LoadAu-thorNet to retrieve detailed information. This sequential dependence is observable across diverse agent tasks (Qin et al. 2023;Ma et al. 2024), confirming that prior tool selections significantly influence subsequent choices. Such inertia is widely present in many domain tasks, where LLM agents are usually applied.

Inspired by these insights, we introduce AutoTool, a novel graph-based method for automatic tool selection in LLM agents. AutoTool constructs a graph representation capturing the observed inertia in tool selection from historical workflows, where nodes correspond to tools and edges encode observed sequential dependencies. Additionally, AutoTool intelligently integrates parameter-level information into this graph, thereby enabling automated parameter filling. By efficiently traversing this structured representation, AutoTool selects the appropriate tools and parameters significantly faster than traditional inference-based approaches. Experimental evaluations demonstrate that AutoTool substantially reduces token consumption and LLM call counts while maintaining comparable agent task progress rates to LLM-based methods.

Our contributions are summarized as follows:

• We empirically identify and analyze the phenomenon of tool usage inertia in LLM-based agents, both in tool selection and parameter filling.

• We design a method to construct an inertia-aware tool graph that captures sequential patterns and data flow in agent behavior.

• We develop a graph-based selection algorithm that effi-arXiv:2511.14650v1 [cs.AI] 18 Nov 2025 ciently determines the next tool and its parameters with minimal LLM intervention. • We conduct extensive experiments showing that AutoTool achieves significant reductions in LLM inference cost while preserving task performance.

2 Background and Related Work

LLM agents have significant potential in solving complex problems, primarily through effective task planning, reasoning, and interaction with external tools (Qin et al. 2024;Qu et al. 2025). The seminal work ReAct (Yao et al. 2023b) introduces the core paradigm of driving agent decisions through interleaved “Thought-Act-Observe” cycles, which has become a cornerstone for numerous open-source frameworks, including Langchain (LangChain 2023) and MetaGPT (Hong et al. 2023). Subsequent research has expanded agents’ capabilities to interact with a vast array of external tools, such as RESTful APIs in RestGPT (Song et al. 2023) or various AI models orchestrated by HuggingGPT (Shen et al. 2023). However, a shared limitation across these powerful frameworks is that the fundamental decision of which tool to use at each step still predominantly relies on a costly LLM inference (Belcak et al. 2025). This reliance creates a significant computational bottleneck, which is the primary issue our work aims to address.

Research in tool selection can be broadly categorized into two types: fine-tuning-dependent and tuning-free methods. Fine-tuning methods like Toolformer (Schick et al. 2023), Gorilla (Patil et al. 2024) and ToolRL (Qian et al. 2025) aim to improve intrinsic tool-use capabilities. While these approaches demonstrably enhance a model’s intrinsic ability to call tools correctly, their reliance on high-quality data or carefully crafted reward signals presents a significant barrier to scalability and adaptability. Tuning-free methods employ different strategies at runtime. Approaches like AnyTool (Du, Wei, and Zhang 2024) and ToolNe

This content is AI-processed based on ArXiv data.