Researchers have proposed LOTUS, a hallucination detector that looks inside a language model rather than judging only its final wording. The method is designed to find signals spread across transformer layers, feature dimensions and tokens without feeding every high-dimensional activation into a large classifier.
LOTUS first converts variable-length token representations into statistical summaries. These descriptors capture peak activations, broader activation distributions and changes between layers. It then applies a low-rank multilinear decomposition, a compression technique that preserves structure across several dimensions, to create a compact representation. A lightweight classifier uses that representation to predict whether an answer is hallucinated.
The authors report better performance than recent baselines across multiple language models and question-answering benchmarks, while reducing dependence on labeled examples. The arXiv abstract does not establish that the detector generalizes to every model, domain or kind of factual error, and the paper is a preprint rather than a deployed safety guarantee. Internal-state methods also require access to model activations, which closed API providers may not expose. The practical contribution is a more structured way to compress those activations without relying only on selected neurons or simple pooling.