AutoTool: Efficient Tool Selection for Large Language Model Agents

Overview

AutoTool is a graph-based and lightweight tool selection framework designed to optimize Large Language Model (LLM) agents.

Instead of passively following observed tool usage inertia, AutoTool manages it as an active behavior. By integrating statistical structures into the agent design, AutoTool leverages usage inertia to effectively address high latency and resource consumption common in existing multi-step tool selection frameworks.

🚀 Installation

1. Setup Conda Environment

First, create the environment and install the core autool package.

conda create --name autotool python=3.10
conda activate autotool
pip install -e .

2. Environment Preparation (AgentBoard)

AutoTool relies on the AgentBoard environment (specifically for Alfworld, ScienceWorld, and ToolQuery-Academic). We highly recommend using Docker.

docker pull zzh202121/agentboard:0117
docker run -itd \
    --network host \
    --name autotool \
    --shm-size 64gb \
    -v /path/to/model_download:/model_download \
    -v /path/to/AutoTool_Repo:/data \
    zzh202121/agentboard:0117 \
    /bin/bash

docker attach autotool

echo 'export PROJECT_PATH=/data' >> ~/.bashrc
source ~/.bashrc
conda activate agentboard

git clone https://github.com/hkust-nlp/AgentBoard.git
cd AgentBoard
mkdir data
wget https://huggingface.co/datasets/hkust-nlp/agentboard/resolve/main/data.tar.gz
tar -zxvf data.tar.gz

3. Download Embedding Model

AutoTool uses SimCSE for embeddings. Download the model to your specified directory:

cd /data/model_download
huggingface-cli download princeton-nlp/sup-simcse-roberta-base \
    --local-dir models--princeton-nlp--sup-simcse-roberta-base/snapshots/4bf73c6b5df517f74188c5e9ec159b2208c89c08 \
    --local-dir-use-symlinks False

⚙️ Configuration

API and Paths

Copy the example configuration and update it with your credentials:

cd autool
cp .env.example .env

Edit .env file:

• Configure API Base URL.
• Configure API Key.
• Set the Embedding Model path.
• Set the Tool Description file path.

📂 Project Structure & Core Components

The project is organized into two main parts: the AutoTool core package and the AgentBoard evaluation framework.

.
├── autool/                         # 📦 Core AutoTool Framework
│   ├── core/
│   │   ├── tool_predict/           # Graph-based tool selection logic
│   │   ├── param_completion/       # Parameter dependency handling & completion
│   └── utils/                      # Utilities for embeddings (SimCSE) and parsing
│
├── agentboard/                     # 🧪 Evaluation Platform (Modified)
│   ├── agents/                     # Agent Implementations (Entry Points)
│   │   ├── test_agent.py           # 👉 AutoTool + ReAct (Main Implementation)
│   │   ├── test_agent2.py          # 👉 Ngram + ReAct (Baseline)
│   ├── prompts/                    # System prompts for Alfworld, ScienceWorld, etc.
│   └── eval_main.py                # 🚀 Main evaluation script
│
└── eval_configs/
    └── main_results_all_tasks.yaml # Global configuration for evaluation

Core Components & Hyperparameters

Key agent implementations are located in agentboard/agents:

• test_agent.py: AutoTool + ReAct
• test_agent2.py: AutoTool + Reflexion

Note: You can adjust hyperparameters at the end of each python file.

To select the agent type for evaluation, modify eval_configs/main_results_all_tasks.yaml.

🏃 Quick Start

Run the evaluation script (e.g., for AlfWorld):

python agentboard/eval_main.py \
    --cfg-path eval_configs/main_results_all_tasks.yaml \
    --tasks alfworld \
    --model DeepSeekV3 \
    --log_path ./results/alfworld_quick_start \
    --project_name evaluate_reflection \
    --baseline_dir ./data/baseline_results

Important Note regarding --model: Although we use API-based inference (not local models), the AgentBoard framework requires the --model argument to be present. You can specify any string (e.g., qwen), but do not omit the flag.

📖 Citation

If you find AutoTool useful for your research, please cite our paper:

@article{jia2025autotool,
  title={AutoTool: Efficient Tool Selection for Large Language Model Agents},
  author={Jia, Jingyi and Li, Qinbin},
  journal={arXiv preprint arXiv:2511.14650},
  year={2025}
}

✉️ Contact

For questions or feedback, feel free to contact: jingyijia@hust.edu.cn