Large language models encode a great deal of factual knowledge, but knowing when to trust a given claim is a separate problem. A new arXiv preprint, Truth in Contrast: Calibrated Factual Evidence from Language Model Likelihoods, takes up this question. The authors note that estimating confidence in factual claims reliably remains challenging, and they investigate likelihood-based confidence from complementary perspectives.

Because the abstract is cut off, the full methodology is not yet visible. The title hints that contrasting likelihoods may be central to the proposed calibration, and the work appears to treat likelihoods as a source of factual evidence rather than simply as a measure of generation probability.

At this stage, the announcement offers only a high-level summary. The paper is listed as arXiv:2610.23527v1, and its abstract indicates a focus on complementary perspectives on likelihood-based confidence. As this is a single source, there is no comparison to make.