Can LLMs Navigate Beliefs and Facts? Depends on How You Phrase It

Humans naturally form and express beliefs in daily communication, e.g., I think the answer is 3'' orI suppose that's right.'' Such beliefs inevitably intertwine with fact and knowledge, making the ability to handle them in tandem desirable for large language models (LLMs), as they are increasingly deployed in user-facing settings. Prior work suggested that even capable LLMs exhibit a systemic weakness in acknowledging user beliefs grounded in incorrect information. We extend this evaluation to 9 LLMs across 18 epistemic expressions and find that the weakness is more nuanced than previously described: it may not exist or exists to a much lesser extent depending on how beliefs are phrased. We further show that the phenomenon stems from task confusion: models default to fact-checking the underlying claim rather than tracking the user's stated belief. This is substantiated by two lines of evidence: (1) an explicit ``do not fact-check'' instruction raises false-claim accuracy from 51\% to 88\%, while a prompt encouraging the model to carry out fact-checking degrades performance, and (2) analyzing chains of thought reveals that models fact-check 38.5\% of the time while being queried for belief confirmation, and collapses accuracy by 49.8\% on false claims. Our findings clarify prior results and illuminate how fact-checking behavior interferes with belief tracking in LLMs.