Faithfulness hallucinations occur when a large language model generates content that does not follow from the provided context, even if it is internally consistent. Detecting these errors is important for high-stakes uses, but existing methods often struggle to separate what the model learned during training from what it is actually reading in the prompt.

The new preprint, posted on arXiv, suggests looking at the unembedding layer—the part of the network that maps internal states to vocabulary probabilities. The authors describe a 'dark space' in this layer that contains signals about how much the model is relying on the current context versus its parametric memory. By contrasting these two modes of knowledge use, they aim to flag outputs that are not grounded in the input.

The paper is a proposal and early analysis rather than a deployed system; the abstract does not include quantitative results. Still, it points to a promising direction: using internal representations, not just output probabilities, to audit model faithfulness. The authors frame this as a step toward safer deployment in domains where unsupported claims are costly.