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

2024-08-08

arXiv.org

AI Summary

Existing LLM tool-use evaluation benchmarks are limited to stateless API calls or offline trajectory analysis, failing to capture realistic multi-turn, state-dependent interactions. Method: We propose ToolSandbox—the first benchmark framework supporting stateful execution, multi-turn interaction, and online dialogue evaluation. It features (1) stateful tool execution with implicit state-dependency modeling; (2) a hybrid rule- and LLM-based user simulator enabling on-policy dynamic dialogue evaluation; and (3) a dynamic criterion mechanism assessing both intermediate and final milestones. Built upon finite-state machine modeling, RESTful sandbox encapsulation, and input normalization, it handles real-world challenges such as incomplete information. Results: Experiments reveal that current SOTA models achieve <40% accuracy on state-dependent tasks, exposing critical limitations in tool orchestration. ToolSandbox establishes a new paradigm and a rigorous, reproducible evaluation standard for LLM tool learning.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Safety and Robustness
• Search and Optimization: Sampling/Simulation-based Search

Application Category

• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

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

Problem

• Evaluating LLM tool-use capabilities comprehensively
• Addressing stateful tool execution and dependencies
• Assessing performance gaps in complex tasks

Innovation

• Stateful tool execution with implicit dependencies
• Built-in user simulator for conversational evaluation
• Dynamic evaluation strategy for milestones