ToolSandbox: A Stateful, Conversational, Interactive Evaluation Benchmark for LLM Tool Use Capabilities

TL;DR

ToolSandbox provides a comprehensive, stateful, conversational benchmark for evaluating large language models' tool use capabilities, revealing significant performance gaps and challenging tasks even for state-of-the-art models.

Contribution

It introduces a novel, dynamic evaluation framework with stateful tool execution and on-policy dialogue support, advancing beyond previous stateless or off-policy assessments.

Findings

Open source and proprietary models show significant performance gaps.

Complex tasks remain challenging for SOTA LLMs.

ToolSandbox offers new insights into LLM tool-use capabilities.

Abstract

Recent large language models (LLMs) advancements sparked a growing research interest in tool assisted LLMs solving real-world challenges, which calls for comprehensive evaluation of tool-use capabilities. While previous works focused on either evaluating over stateless web services (RESTful API), based on a single turn user prompt, or an off-policy dialog trajectory, ToolSandbox includes stateful tool execution, implicit state dependencies between tools, a built-in user simulator supporting on-policy conversational evaluation and a dynamic evaluation strategy for intermediate and final milestones over an arbitrary trajectory. We show that open source and proprietary models have a significant performance gap, and complex tasks like State Dependency, Canonicalization and Insufficient Information defined in ToolSandbox are challenging even the most capable SOTA LLMs, providing brand-new insights into tool-use LLM capabilities. ToolSandbox evaluation framework is released at https://github.com/apple/ToolSandbox