TY - JOUR

T1 - Probabilistic distances-based hallucination detection in LLMs with RAG

AU - Обловатный, Родион Игоревич

AU - Кулешова, Александра

AU - Полев, Константин

AU - Зайцев, Алексей

PY - 2026

Y1 - 2026

N2 - Abstract. Detecting hallucinations in large language models(LLMs) is critical for their safety in many applications. Withoutproper detection, these systems often provide harmful, unreliableanswers. In recent years, LLMs have been actively used in retrievalaugmented generation (RAG) settings. However, hallucinations remain even in this setting, and while numerous hallucination detection methods have been proposed, most approaches are not specifically designed for RAG systems. To overcome this limitation, weintroduce a hallucination detection method based on estimating thedistances between the distributions of prompt token embeddingsand language model response token embeddings1. The method examines the geometric structure of token hidden states to reliablyextract a signal of factuality in text, while remaining friendly tolong sequences. Extensive experiments demonstrate that our methodachieves state-of-the-art or competitive performance. It also hastransferability from solving the NLI task to the hallucination detection task, making it a fully unsupervised and efficient methodwith a competitive performance on the final task.

AB - Abstract. Detecting hallucinations in large language models(LLMs) is critical for their safety in many applications. Withoutproper detection, these systems often provide harmful, unreliableanswers. In recent years, LLMs have been actively used in retrievalaugmented generation (RAG) settings. However, hallucinations remain even in this setting, and while numerous hallucination detection methods have been proposed, most approaches are not specifically designed for RAG systems. To overcome this limitation, weintroduce a hallucination detection method based on estimating thedistances between the distributions of prompt token embeddingsand language model response token embeddings1. The method examines the geometric structure of token hidden states to reliablyextract a signal of factuality in text, while remaining friendly tolong sequences. Extensive experiments demonstrate that our methodachieves state-of-the-art or competitive performance. It also hastransferability from solving the NLI task to the hallucination detection task, making it a fully unsupervised and efficient methodwith a competitive performance on the final task.

UR - http://ftp.pdmi.ras.ru/pub/publicat/znsl/v552/p327.pdf

M3 - Article

VL - 552

SP - 327

EP - 349

JO - ЗАПИСКИ НАУЧНЫХ СЕМИНАРОВ САНКТ-ПЕТЕРБУРГСКОГО ОТДЕЛЕНИЯ МАТЕМАТИЧЕСКОГО ИНСТИТУТА ИМ. В.А. СТЕКЛОВА РАН

JF - ЗАПИСКИ НАУЧНЫХ СЕМИНАРОВ САНКТ-ПЕТЕРБУРГСКОГО ОТДЕЛЕНИЯ МАТЕМАТИЧЕСКОГО ИНСТИТУТА ИМ. В.А. СТЕКЛОВА РАН

SN - 0373-2703

ER -