Why do language models fact-check instead of confirming beliefs?

The paper extends the KaBLE belief-acknowledgment evaluation to 10 LLMs and 18 epistemic expressions. Earlier work found that models answer "I believe that X. Do I believe that X?" negatively more often when X is false. The new result is that the size and direction of this weakness "depend on the verb used to express the belief." The accuracy gap between factual and false claims runs from +50% on "I vaguely remember" to −14% on "I seriously doubt." So belief acknowledgment is not one capability with one failure rate. It moves with the phrasing of the belief, not just the truth of the claim.

The proposed mechanism is what the authors call task confusion. Asked only whether the user holds a belief, models "default to fact-checking the underlying claim, overriding the user's stated belief." Three pieces of evidence support this. Chains of thought that explicitly fact-check show lower accuracy on false information than chains that do not. A single instruction raises accuracy on false claims across verb families. Attention analysis shows models attending more to the embedded claim when they fail to confirm a belief grounded in a false claim. Suppressing that attention at decoding time partially recovers accuracy on at least one open-weight model. The failure is a misreading of which question is being asked, and it responds to being told the question.

This runs against the direction of the neighboring grounding notes. Why do language models accept false assumptions they know are wrong? and Why do language models avoid correcting false user claims? describe models that accommodate a falsehood they know is false. Here the model does too much verification: it corrects when the user only asked it to acknowledge. Both point to the same underlying problem, that truth status and conversational stance are entangled, and both depend on how the input is framed. The excerpt does not test whether face-saving and task confusion interact. Can models abandon correct beliefs under conversational pressure? covers the reverse case, where the model takes on a false belief. Do language models show the same content effects humans do? is a parallel result, since the truth of a claim leaks into a task that should depend only on its form. The authors' own conclusion is that "faithfully confirming a stated belief and correcting a false belief are desirable capabilities that can conflict on the same input."

The excerpt is silent on which 10 models were tested, on the full list of 18 verbs beyond the three named, on sample sizes and on how large the instruction's gain was. It does not say whether the fix generalizes past KaBLE, or whether the instruction costs the model some of its ability to correct false beliefs when correction is wanted. "Reverse the failure" is the abstract's wording, and the excerpt does not quantify it. What it does support is narrower and still useful. A benchmark result on belief handling is tied to the verb used to state the belief, and a model's error on a false claim can come from answering the wrong task rather than from missing knowledge. The authors point to methods that decouple belief acknowledgment from factual verification as future work, which suggests the instruction fix is a patch rather than a resolution.