New Criterion FEPoID for Automatic Layer Selection in LLM Hallucination Detection

A new study on arXiv (2605.26366) proposes First Effective Peak of Intrinsic Dimension (FEPoID) as a criterion for automatically selecting intermediate layers in large language models (LLMs) for hallucination detection. Previous research showed hallucination signals are stronger in intermediate layers than the final layer, but automated selection methods were lacking. The authors tested multiple hypotheses across diverse LLM architectures, scales, and tasks including question answering and summarization benchmarks, finding none consistently effective. FEPoID aims to fill this gap by identifying the first effective peak of intrinsic dimension. The study is authored by researchers from an undisclosed institution and was published on arXiv.