Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Ngu, Noel, Lee, Nathaniel, Shakarian, Paulo
Format: Preprint
Veröffentlicht: 2023
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2308.11189
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866913212065120256
author Ngu, Noel
Lee, Nathaniel
Shakarian, Paulo
author_facet Ngu, Noel
Lee, Nathaniel
Shakarian, Paulo
contents Error prediction in large language models often relies on domain-specific information. In this paper, we present measures for quantification of error in the response of a large language model based on the diversity of responses to a given prompt - hence independent of the underlying application. We describe how three such measures - based on entropy, Gini impurity, and centroid distance - can be employed. We perform a suite of experiments on multiple datasets and temperature settings to demonstrate that these measures strongly correlate with the probability of failure. Additionally, we present empirical results demonstrating how these measures can be applied to few-shot prompting, chain-of-thought reasoning, and error detection.
format Preprint
id arxiv_https___arxiv_org_abs_2308_11189
institution arXiv
publishDate 2023
record_format arxiv
spellingShingle Diversity Measures: Domain-Independent Proxies for Failure in Language Model Queries
Ngu, Noel
Lee, Nathaniel
Shakarian, Paulo
Computation and Language
Artificial Intelligence
Machine Learning
Error prediction in large language models often relies on domain-specific information. In this paper, we present measures for quantification of error in the response of a large language model based on the diversity of responses to a given prompt - hence independent of the underlying application. We describe how three such measures - based on entropy, Gini impurity, and centroid distance - can be employed. We perform a suite of experiments on multiple datasets and temperature settings to demonstrate that these measures strongly correlate with the probability of failure. Additionally, we present empirical results demonstrating how these measures can be applied to few-shot prompting, chain-of-thought reasoning, and error detection.
title Diversity Measures: Domain-Independent Proxies for Failure in Language Model Queries
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2308.11189