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

Quang Minh Nguyen, Luis Frentzen Salim

Keywords: belief, fact-checking, llm, uncertainty

TL;DR: LLMs can navigate beliefs and facts better than expected, depending on the epistemic expression you use; prior observed failure modes can be attributed to task confusion, shown through prompting intervention and chain-of-thought analysis

Abstract: Humans naturally form and express beliefs in daily communication, e.g., "I think the answer is 3'' or "I 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 showed that even capable LLMs exhibit a systemic weakness in acknowledging user beliefs grounded in incorrect information. We extend this evaluation to 11 LLMs across 18 epistemic expressions and find that the size and direction of the weakness depend on the verb used to express the belief: the factual-vs-false accuracy gap ranges from $+52\%$ ("I vaguely remember'') to $-13\%$ ("I seriously doubt''). 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 on the positive-belief subset from 51\% to 83\%, while a prompt encouraging the model to carry out fact-checking degrades performance further, and (2) analyzing chains of thought reveals that 40.9\% of responses explicitly fact-check the underlying claim, with false-claim accuracy of 24.4\% on these versus 78.1\% on the rest. Our findings clarify prior results and illuminate how fact-checking, an otherwise desirable behavior, can interfere with belief tracking in LLMs.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 90