Research map / Multi-agent systems and games / LLM agents

AutoTool: Efficient Tool Selection for Large Language Model Agents

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes AutoTool, a training-free graph-based tool selection framework that builds a Tool Inertia Graph from historical agent trajectories to bypass LLM calls for predictable tool choices and parameter filling.

Problem

High inference cost of LLM-based tool selection in agent loops such as ReAct, which calls the LLM at every step to choose the next tool and its parameters.

Approach

Exploits "tool usage inertia" (sequential regularity in tool invocations) by constructing a directed tool-transition graph with parameter-dependency edges from past trajectories, scoring candidate next tools with a combined frequency and semantic-context score (CIPS), and filling parameters by backtracking the graph; the LLM is invoked only when the inertial attempt fails.

Key results

• On AlfWorld, ReAct+AutoTool raises progress rate from 0.394 to 0.531 while cutting input tokens by 1.60x and LLM calls by 1.18x.
• On ScienceWorld, ReAct+AutoTool keeps progress rate roughly unchanged (0.716 to 0.708) with 1.31x fewer LLM calls and 1.30x fewer input tokens.
• Reflexion+AutoTool on Tool-Query-Academia gives progress rate 0.923 vs 0.917 with 1.26x fewer LLM calls.

Surprising findings

• On AlfWorld, ReAct+AutoTool improves progress rate substantially, which the authors attribute mainly to a fault-tolerance recovery path rather than the graph alone.
• Under a sensitivity sweep, a more lenient inertia threshold did not hurt progress rate, attributed to the 30% inertia-call cap and no-consecutive-inertia constraint.

Limitations

• Parameter filling can select incorrect context (e.g., ignoring current location), causing action failures.
• Inertia can cause redundant actions that waste steps when state already satisfies the goal.
• Overgeneralization of tool patterns to objects that do not support the action.

Experimental setup

Datasets: AlfWorld ScienceWorld ToolQuery-Academic (AgentBoard) ToolBench G3

Benchmarks: AlfWorld ScienceWorld ToolQuery-Academic

Metrics: progress rate LLM calls input tokens output tokens speedup

Baselines: ReAct Reflexion

Builds on: ReAct Reflexion AgentBoard

Closest papers by meaning

• AutoTool: Dynamic Tool Selection and Integration for Agentic Reasoning (ICML 2026 · similarity 0.95)
• ToolSearcher: Optimizing Tool Selection at Scale via Reinforcement Learning (NeurIPS 2026 · similarity 0.91)
• Joint Optimization of Tool Creation and Use for Large Language Model Agents (NeurIPS 2026 · similarity 0.90)
• TaskBench: Benchmarking Large Language Models for Task Automation (NeurIPS 2024 · similarity 0.90)
• MetaTool Benchmark for Large Language Models: Deciding Whether to Use Tools and Which to Use (ICLR 2024 · similarity 0.89)
• AgentTTS: Large Language Model Agent for Test-time Compute-optimal Scaling Strategy in Complex Tasks (NeurIPS 2025 · similarity 0.89)