A new preprint proposes catching AI hallucinations by watching a model's own confidence scores over time, instead of digging into its hidden layers.
The system, called HALT (Hallucination Assessment via Log-probs as Time series), is described in an arXiv preprint (arXiv:2602.02888, https://arxiv.org/abs/2602.02888). It feeds only the top-20 token log-probabilities from a model's output into a small gated recurrent unit, treating those probabilities as a time series and adding entropy-based features on top. That sidesteps both common approaches: it doesn't need a model's internal weights or attention maps like white-box detectors do, and it doesn't read the generated text itself like black-box detectors do, so it can run against proprietary models available only through an API. The authors also built a companion benchmark, HUB, that folds ten task types - among them math, code generation, summarization, and world knowledge - into one evaluation suite.
On that benchmark, HALT beat Lettuce, a fine-tuned ModernBERT-based hallucination detector, while being 30 times smaller and running 60 times faster. The preprint does not report the accuracy margin behind that "outperforms" claim - only the size and speed numbers are quantified. That distinction matters for anyone weighing whether to trust it: a detector cheap enough to run on every generation is a different proposition from one proven to catch meaningfully more hallucinations.
It's also not peer-reviewed yet, and a comparison against a single rival encoder is a thin basis for "outperforms" until the accuracy numbers show up.