Semantic Reformulation Entropy for Robust Hallucination Detection in QA Tasks

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Tong, Chaodong, Zhang, Qi, Jiang, Lei, Liu, Yanbing, Sun, Nannan, Li, Wei
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866909806399324160
author Tong, Chaodong
Zhang, Qi
Jiang, Lei
Liu, Yanbing
Sun, Nannan
Li, Wei
author_facet Tong, Chaodong
Zhang, Qi
Jiang, Lei
Liu, Yanbing
Sun, Nannan
Li, Wei
contents Reliable question answering with large language models (LLMs) is challenged by hallucinations, fluent but factually incorrect outputs arising from epistemic uncertainty. Existing entropy-based semantic-level uncertainty estimation methods are limited by sampling noise and unstable clustering of variable-length answers. We propose Semantic Reformulation Entropy (SRE), which improves uncertainty estimation in two ways. First, input-side semantic reformulations produce faithful paraphrases, expand the estimation space, and reduce biases from superficial decoder tendencies. Second, progressive, energy-based hybrid clustering stabilizes semantic grouping. Experiments on SQuAD and TriviaQA show that SRE outperforms strong baselines, providing more robust and generalizable hallucination detection. These results demonstrate that combining input diversification with multi-signal clustering substantially enhances semantic-level uncertainty estimation.
format Preprint
id arxiv_https___arxiv_org_abs_2509_17445
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Semantic Reformulation Entropy for Robust Hallucination Detection in QA Tasks
Tong, Chaodong
Zhang, Qi
Jiang, Lei
Liu, Yanbing
Sun, Nannan
Li, Wei
Computation and Language
Reliable question answering with large language models (LLMs) is challenged by hallucinations, fluent but factually incorrect outputs arising from epistemic uncertainty. Existing entropy-based semantic-level uncertainty estimation methods are limited by sampling noise and unstable clustering of variable-length answers. We propose Semantic Reformulation Entropy (SRE), which improves uncertainty estimation in two ways. First, input-side semantic reformulations produce faithful paraphrases, expand the estimation space, and reduce biases from superficial decoder tendencies. Second, progressive, energy-based hybrid clustering stabilizes semantic grouping. Experiments on SQuAD and TriviaQA show that SRE outperforms strong baselines, providing more robust and generalizable hallucination detection. These results demonstrate that combining input diversification with multi-signal clustering substantially enhances semantic-level uncertainty estimation.
title Semantic Reformulation Entropy for Robust Hallucination Detection in QA Tasks
topic Computation and Language
url https://arxiv.org/abs/2509.17445