When2Tool

Benchmark dataset for "LLM Agents Already Know When to Call Tools — Even Without Reasoning" (arXiv:2605.09252).

Overview

When2Tool is a benchmark of 18 environments designed to study when LLM agents should call tools. Tasks range from trivially solvable without tools to impossible without them, across three categories of tool necessity:

• Computational Scale (5 envs): Calculator, Statistics, Counting, Matrix, Prime
• Knowledge Boundaries (5 envs): Retriever, Historical Year, Game Rule, Hash, Decoding
• Execution Reliability (5 envs): List Manipulation, DateTime, Code Executor, Schedule, Regex Match
• Multi-hop (3 envs): Calculator, Retriever, Code Executor (3-step chains)

Each environment has three difficulty levels (easy, medium, hard) that create a clear decision boundary between tool-necessary and tool-unnecessary tasks.

Dataset Structure

Configs

• single_hop: 15 single-hop environments (900 train / 2,250 test)
• multi_hop: 3 multi-hop environments with 3-step chains (180 train / 450 test)

Fields

Field	Type	Description
id	int	Unique task identifier
difficulty	str	easy, medium, or hard
multi_step	bool	Whether the task requires multiple tool calls
instruction	str	The task instruction given to the agent
env_name	str	Environment name (e.g., CalculatorEnv)
tools	str (JSON)	Available tools for this environment
parameters	str (JSON)	Environment parameters (e.g., corpus for retriever)
answer	str	Expected final answer
steps	str (JSON)	Intermediate steps for multi-hop tasks
tags	str (JSON)	Environment and task type tags

Loading

from datasets import load_dataset

# Single-hop tasks
ds = load_dataset("Trustworthy-ML-Lab/When2Tool", "single_hop")

# Multi-hop tasks
ds_mh = load_dataset("Trustworthy-ML-Lab/When2Tool", "multi_hop")

# Access a sample
print(ds["test"][0]["instruction"])
print(ds["test"][0]["env_name"])

Citation

@article{sun2026when2tool,
  title={LLM Agents Already Know When to Call Tools -- Even Without Reasoning},
  author={Sun, Chung-En and Liu, Linbo and Yan, Ge and Wang, Zimo and Weng, Tsui-Wei},
  journal={arXiv preprint arXiv:2605.09252},
  year={2026},
  url={https://arxiv.org/abs/2605.09252}
}