A new preprint on arXiv introduces a behavioral probe called first-token cueing to investigate why large language models give wrong answers to factual questions in closed-book settings. The authors note that such errors are typically blamed on missing knowledge, but they present evidence that models often encode the correct answers even when they fail to produce them.

The abstract suggests that many of these errors may be recall failures rather than knowledge gaps. By cueing the first token of a response, the method appears to reveal whether the model actually has the information stored internally but struggles to retrieve it spontaneously.

The paper is available as arXiv:2610.23400. Since only the abstract was provided, the full details of the experimental setup and results are not yet accessible, but the core claim—that some errors are misattributed to missing knowledge—offers a new angle for diagnosing and potentially improving LLM factual accuracy.