In this study, we investigated how the complexity of a query can influence tool-use hallucinations in large language models that were tasked with setting an appointment. Using explainable AI (XAI) as a tool to test the hypothesis, we analyzed trends in hallucination rates by the type of query and saw results that when complexity and ambiguity of the query increased (such as asking for multiple days or unclear time ranges), the model was more likely to make errors due to misunderstanding time-related data from tool outputs. This shows that making large language models (LLMs) understand what seems logical to a human is still difficult, especially in complex tasks, and that knowing what causes these errors can help build more trustworthy AI systems.
Read More...