Tool-augmented language models are meant to improve reliability by grounding responses in external tool outputs, moving beyond what parametric knowledge alone can supply. But grounding is not a guarantee of correctness. The abstract for this paper notes a stubborn failure mode: even when the correct tool is called and returns accurate information, the model may still produce a hallucinated response.

To address this, the authors introduce ToolHACE, a dataset built for detecting hallucinations in tool-based LLM responses. The focus is not on wrong tool selection or faulty tool output, but on the model's own output going wrong despite receiving good input from the tool.

The available abstract cuts off mid-sentence, so this summary cannot speak to the dataset's size, construction methodology, or any reported results. What is clear from the source is the motivation: a resource aimed at a failure mode that standard tool-augmentation evaluation may overlook.