Guardado en:
Detalles Bibliográficos
Autores principales: Moore, Kyle, Roberts, Jesse, Watson, Daryl, Wisniewski, Pamela
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:https://arxiv.org/abs/2503.12528
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866916653794590720
author Moore, Kyle
Roberts, Jesse
Watson, Daryl
Wisniewski, Pamela
author_facet Moore, Kyle
Roberts, Jesse
Watson, Daryl
Wisniewski, Pamela
contents Recent work has sought to quantify large language model uncertainty to facilitate model control and modulate user trust. Previous works focus on measures of uncertainty that are theoretically grounded or reflect the average overt behavior of the model. In this work, we investigate a variety of uncertainty measures, in order to identify measures that correlate with human group-level uncertainty. We find that Bayesian measures and a variation on entropy measures, top-k entropy, tend to agree with human behavior as a function of model size. We find that some strong measures decrease in human-similarity with model size, but, by multiple linear regression, we find that combining multiple uncertainty measures provide comparable human-alignment with reduced size-dependency.
format Preprint
id arxiv_https___arxiv_org_abs_2503_12528
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Investigating Human-Aligned Large Language Model Uncertainty
Moore, Kyle
Roberts, Jesse
Watson, Daryl
Wisniewski, Pamela
Computation and Language
Recent work has sought to quantify large language model uncertainty to facilitate model control and modulate user trust. Previous works focus on measures of uncertainty that are theoretically grounded or reflect the average overt behavior of the model. In this work, we investigate a variety of uncertainty measures, in order to identify measures that correlate with human group-level uncertainty. We find that Bayesian measures and a variation on entropy measures, top-k entropy, tend to agree with human behavior as a function of model size. We find that some strong measures decrease in human-similarity with model size, but, by multiple linear regression, we find that combining multiple uncertainty measures provide comparable human-alignment with reduced size-dependency.
title Investigating Human-Aligned Large Language Model Uncertainty
topic Computation and Language
url https://arxiv.org/abs/2503.12528