Guardado en:
Detalles Bibliográficos
Autores principales: Yoon, WonJin, Bulovic, Ian, Miller, Timothy A.
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:https://arxiv.org/abs/2502.15018
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866914166769451008
author Yoon, WonJin
Bulovic, Ian
Miller, Timothy A.
author_facet Yoon, WonJin
Bulovic, Ian
Miller, Timothy A.
contents Large language models perform surprisingly well on many zero-shot classification tasks, but are difficult to fairly compare to supervised classifiers due to the lack of a modifiable decision boundary. In this work, we propose and evaluate a method that transforms binary classification tasks into pairwise comparisons between instances within a dataset, using LLMs to produce relative rankings of those instances. Repeated pairwise comparisons can be used to score instances using the Elo rating system (used in chess and other competitions), inducing a confidence ordering over instances in a dataset. We evaluate scheduling algorithms for their ability to minimize comparisons, and show that our proposed algorithm leads to improved classification performance, while also providing more information than traditional zero-shot classification.
format Preprint
id arxiv_https___arxiv_org_abs_2502_15018
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Using tournaments to calculate AUROC for zero-shot classification with LLMs
Yoon, WonJin
Bulovic, Ian
Miller, Timothy A.
Computation and Language
Large language models perform surprisingly well on many zero-shot classification tasks, but are difficult to fairly compare to supervised classifiers due to the lack of a modifiable decision boundary. In this work, we propose and evaluate a method that transforms binary classification tasks into pairwise comparisons between instances within a dataset, using LLMs to produce relative rankings of those instances. Repeated pairwise comparisons can be used to score instances using the Elo rating system (used in chess and other competitions), inducing a confidence ordering over instances in a dataset. We evaluate scheduling algorithms for their ability to minimize comparisons, and show that our proposed algorithm leads to improved classification performance, while also providing more information than traditional zero-shot classification.
title Using tournaments to calculate AUROC for zero-shot classification with LLMs
topic Computation and Language
url https://arxiv.org/abs/2502.15018