A new preprint from arXiv highlights a stubborn problem in long-context AI: large language models frequently attempt to answer questions even when the provided material contains no answer. The authors argue that this failure to recognize unanswerable questions is a critical reliability gap, especially in information-seeking tasks where users expect accuracy over confident fabrication.
The paper introduces FactGuard, a framework aimed at detecting unanswerable questions before the model commits to a response. While the abstract does not provide evaluation details or performance figures, the proposal positions FactGuard as a guardrail for long-context reasoning, where models must weigh evidence across many passages.
Because this is a single preprint, the claims are preliminary and have not been peer-reviewed. Still, the work points to a practical direction: making LLMs more trustworthy by teaching them to abstain when the evidence is insufficient.