Researchers have found that large language models (LLMs) struggle to navigate beliefs and facts when they are phrased in certain ways. The study evaluated 10 LLMs across 18 different phrases used to express beliefs, such as 'I think' or 'I suppose.' The results showed that the accuracy of these models depends on the verb used to express the belief, with some verbs leading to a +50% increase in accuracy and others resulting in a -14% decrease. This phenomenon is attributed to task confusion, where the model defaults to fact-checking rather than tracking the user's stated belief. The study suggests that this issue can be addressed through future research on intervention methods. --- Why it matters: This matters because LLMs are increasingly being used in user-facing settings, such as chatbots and virtual assistants, where they need to handle users' beliefs and facts accurately. If these models struggle with phrasing, it could lead to miscommunication and incorrect information being conveyed to users. Source: https://arxiv.org/abs/2608.17809