Benchmarking LLM Tool-Use in the Wild

1. Problem Statement

Current Large Language Model (LLM) tool-use benchmarks suffer from a significant gap between evaluation environments and real-world user interactions. Existing benchmarks (e.g., BFCL, ToolBench, $\tau$-Bench) are often overly idealized, featuring:

• Isolated Tasks: Tasks are independent and lack semantic correlation across turns.
• Explicit Intent: User instructions are complete, unambiguous, and provide all necessary information upfront.
• Static Policies: Users rarely switch between task types (e.g., from tool-use to casual chat) within a single session.

Consequently, LLMs appear to have mastered tool-use on these benchmarks, but they fail when faced with "wild" user behaviors characterized by compositional tasks, implicit intents, and frequent instruction transitions. This leads to spurious progress in agentic capabilities.

2. Methodology: WildToolBench

The authors introduce WildToolBench, a benchmark grounded in real-world user behavior patterns derived from large-scale analysis of user logs.

A. Core Design Philosophy

WildToolBench focuses on three specific challenges inherent to real user interactions:

1. Compositional Tasks: Users often bundle multiple simple requirements into a single instruction (e.g., "Check weather for Chicago and LA, then find a hotel"). This requires the LLM to orchestrate complex tool-call topologies (trees/parallel execution) rather than simple chains.
2. Implicit Intent (Hidden Information): User intent is often spread across dialogue turns. The LLM must infer missing information via:
   • Partial Information: Omitting details present in previous turns.
   • Coreferential Reference: Using pronouns or ellipsis referring to past entities.
   • Long-Range Dependency: Requiring context from distant turns ($i-j > 2$).
3. Instruction Transition: Users naturally switch between modes (task execution, clarification, casual chat, follow-up) within a session. The LLM must dynamically adapt its policy (e.g., from tool-calling to direct answering or clarification).

B. Data Curation Pipeline

The dataset was constructed through a rigorous human-in-the-loop process:

1. Seed Extraction: Analyzed real user logs to identify behavior patterns and seed scenarios.
2. Tool Set: Curated ~1,600 APIs from public repositories, filtered down to 400 tool lists covering diverse domains (Weather, Finance, Media, etc.).
3. Multi-Agent Simulation: Used high-performance LLMs to simulate User and Assistant agents, generating initial trajectories based on the three challenges.
4. Human Verification: 9 human experts manually annotated and refined 256 scenarios (1,024 tasks total). They verified tool dependencies, corrected hallucinations, and ensured natural dialogue flow (e.g., rewriting synthetic "robotic" dialogue to be more colloquial).
5. Evaluation Metrics:
   • Session Accuracy: Whether all tasks in a dialogue are completed correctly.
   • Task Accuracy: Accuracy per individual task.
   • Orchestration Metrics:
     • Optimal Path Rate (OP): Percentage of times the LLM chose the most efficient tool execution path (shortest depth).
     • Accomplish Progress Rate (AP): Percentage of successful nodes in the tool execution tree.

3. Key Contributions

• WildToolBench Dataset: A high-quality benchmark of 1,024 tasks across 256 multi-turn sessions, explicitly modeling compositional, implicit, and transitional user behaviors.
• Novel Evaluation Framework: Introduced fine-grained metrics (OP Rate, AP Rate) to evaluate tool orchestration efficiency, moving beyond simple binary success/failure.
• Comprehensive Benchmarking: Evaluated 57 LLMs, including proprietary models (GPT-4o, Claude-4, Gemini-2.5), open-source general models (Qwen, DeepSeek, GLM), and specialized tool-use models.

4. Experimental Results

The results reveal a stark reality: Current LLMs struggle significantly in "wild" settings.

• Low Overall Performance: No model achieved a session accuracy higher than 15%. The best-performing model (Gemini-2.0-Thinking) reached only 14.45% session accuracy. Most models scored below 60% on individual task accuracy.
• Tool Orchestration Failure:
  • LLMs perform poorly on mixed sequential/parallel tasks ($g^{S+P}_{multi}$), with accuracy dropping to ~25%.
  • Optimal Path (OP) rates are low (peak ~42%), indicating models often choose inefficient execution paths.
  • Specialized tool-use models (e.g., xLAM, Hammer) often performed worse than general-purpose models, suggesting poor generalization.
• Hidden Intent Challenges:
  • Long-Range Dependency is the hardest challenge (accuracy < 50% for all models).
  • Reasoning models (e.g., o1, DeepSeek-R1) showed better performance on partial information but still struggled with long-range context.
• Instruction Transition Impact: As the frequency of instruction transitions (switching between chat, clarify, and tool-use) increases, model accuracy drops significantly (up to 30% decrease), attributed to "self-conditioning" where previous actions bias subsequent decisions.
• Error Analysis:
  • The primary failure mode has shifted from syntactic errors to semantic/logical reasoning.
  • "Wrong Name" (selecting the wrong tool) and "Missing Info" (failing to ask for clarification) are the most prevalent errors.
  • Models exhibit two failure profiles: "Cautious" (high refusal rates) vs. "Eager" (high hallucination/wrong tool rates).

5. Significance and Conclusion

• Redefining Agentic Evaluation: The paper argues that the bottleneck for LLM agents is no longer just tool invocation syntax, but understanding user behavior. Effective agents must possess "Theory of Mind" to handle vagueness, compositionality, and dynamic context switching.
• Limitations of Current Benchmarks: Existing benchmarks are saturated and fail to capture the complexity of real-world interactions, leading to an overestimation of LLM capabilities.
• Future Directions: The authors emphasize that future progress requires models to improve in long-context comprehension, instruction following under ambiguity, and dynamic policy switching, rather than just scaling up reasoning capabilities for static tasks.

In summary, WildToolBench exposes a substantial gap in the robustness of current LLMs' agentic abilities, demonstrating that while models can solve isolated, idealized tasks, they are currently ill-equipped to handle the messy, flexible, and context-dependent nature of real human interaction.