Benchmarking LLM Tool-Use in the Wild
Summary: arXiv:2604.06185v1 Announce Type: cross
Abstract
Fulfilling user needs through Large Language Model (LLM) multi-turn, multi-step tool-use is rarely a straightforward process. Real user interactions are inherently wild, being intricate, messy, and flexible. In this article, we identify three key challenges emerging from user behavior:
- Compositional Tasks: These tasks demand efficient orchestration of tool-call topologies, presenting a challenge for LLMs to manage complex interactions.
- Implicit Intent: User intent is often spread across dialogue turns, necessitating contextual inference for accurate understanding and response.
- Instruction Transition: The mixing of task queries, clarifications, and casual conversation forces LLMs to adjust their policies on the fly, complicating the interaction further.
Current benchmarks often overlook these behaviors, leading to an inflated perception of the progress made by LLMs in tool-use capabilities. To address this gap, we introduce WildToolBench, a new benchmark specifically designed for LLM tool-use that is grounded in real-world user behavior patterns.
Introduction to WildToolBench
WildToolBench aims to provide a more accurate reflection of LLM performance in real-world scenarios. By focusing on the intricacies of user interactions, it enables a better understanding of how LLMs can effectively utilize tools in a way that meets user needs.
Key Findings
Through comprehensive evaluations involving 57 different LLMs, we discovered that none of the models achieved an accuracy rate higher than 15%. This finding highlights a significant gap in the robustness of LLMs’ agentic abilities in practical applications.
Controlled Experiments and Analysis
Further controlled experiments and in-depth analyses have illuminated that the primary challenge for LLM tool-use does not stem from tasks that are artificially complex. Rather, it arises from the chaotic and unpredictable nature of user behavior. This underscores the necessity of re-evaluating how LLMs interact with users and tools.
Conclusion
As the field of artificial intelligence continues to advance, it is imperative to rethink the frameworks we use for benchmarking LLM tool-use. WildToolBench serves as a critical step toward developing more robust and user-aligned AI systems. By understanding the real-world complexities of user interactions, researchers and developers can work towards enhancing the capabilities of LLMs, ultimately leading to more effective and reliable tool-use in diverse applications.