ToolFailBench: Diagnosing Tool-Use Failures in LLM Agents

2026-07-06

AI Summary

Existing evaluation frameworks struggle to discern fine-grained failure modes of large language models in tool use—such as skipping tools, ignoring their outputs, or invoking them unnecessarily—often leading to misleading capability assessments. To address this, this work introduces ToolFailBench, a diagnostic benchmark comprising 1,000 tasks spanning finance, medicine, and law, which systematically defines and annotates four distinct tool-use failure patterns. By contrasting tasks that strictly require tool invocation with those where tool use is prohibited, the benchmark reveals substantial differences in failure mechanisms among models that otherwise achieve similar overall scores. Annotations are generated via majority voting among a rule-based classifier and two large language model judges. Evaluation across 19 prominent models shows that the best-performing model achieves an 86.33% Clean Tool-Use rate, while Llama-3.1 variants exhibit a strong tendency toward over-invocation, and models of comparable scale differ by up to 89 percentage points in control-task accuracy.

Abstract

Tool calling is central to modern language model agents, but aggregate benchmark scores often hide where tool use fails. A model that never calls a needed tool and a model that calls the tool but ignores the result can look similar under final task accuracy. We introduce ToolFailBench, a diagnostic benchmark for measuring tool-use failures across 1,000 tasks in finance, medicine, law, cybersecurity, and real estate. Tool-required tasks return values the model wouldn't guess, forcing it to trust the tool while control tasks attach the same tools but should be answered directly. We label each trace with Tool-Skip, Result-Ignore, Output-Fabrication, and Unnecessary-Tool-Use, using a rule classifier and two LLM judges aggregated by majority vote. Across 19 headline models, the best reaches 86.33% Clean Tool-Use Rate, showing that faithful tool use is not saturated. More importantly, models with similar aggregate scores fail in different ways: most stay disciplined on no-tool controls, while Llama-3.1 models show an Always-Call pattern, and at the same parameter scale Llama-3.1-70B and Qwen2.5-72B differ by 89 percentage points on control-task accuracy. Tool-use evaluation should measure not only whether agents call tools, but whether they use tool outputs correctly and avoid tools when none is needed.

Problem

• tool-use failures
• LLM agents
• diagnostic benchmark
• tool calling
• output fabrication

Innovation

• tool-use failure diagnosis
• diagnostic benchmark
• LLM agent evaluation
• clean tool-use rate
• failure mode analysis