Researchers have developed a new method for detecting token-level hallucinations in generated text. This approach treats hallucination as a temporally extended span and uses sequence labeling to detect it. The method fuses multiple signals, including text statistics, Natural Language Inference entailment, and language model surprisal, to score each token. Experiments show an 11-point gain over a baseline method, with the detector working on closed-source models and maintaining performance even when applied to new text produced by unseen language models. --- Why it matters: This matters because current methods for detecting hallucinations can fail when the generating model is confident in its output. This new approach shows promise in addressing this limitation and could be useful for improving the reliability of generated text in applications such as chatbots, virtual assistants, and language translation systems. Source: https://arxiv.org/abs/2608.18115