Detects hallucinations at test time by tracking usable-information deficiencies across model layers—especially under ambiguous or unanswerable prompts—without extra training or architecture changes.
Overview
Hallucinations are hard to catch when a model sounds fluent. This paper looks inside the network and asks whether usable information is missing as representations move through layers.
We propose a layer-wise information deficiency signal that flags hallucinations at test time, with particular strength on ambiguous or unanswerable prompts. The method needs no additional training and leaves the model architecture unchanged.
Why it matters
A practical detector should work on frozen models and under realistic uncertainty. By reading information loss across layers, this approach offers a training-free check that complements decoding-time safeguards.
Paper
The full paper is embedded below. You can also download the PDF or view it on arXiv.