Saved in:
Bibliographic Details
Main Authors: Zong, Yongshuo, Yu, Tingyang, Chavhan, Ruchika, Zhao, Bingchen, Hospedales, Timothy
Format: Preprint
Published: 2023
Subjects:
Online Access:https://arxiv.org/abs/2310.01651
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866929446193201152
author Zong, Yongshuo
Yu, Tingyang
Chavhan, Ruchika
Zhao, Bingchen
Hospedales, Timothy
author_facet Zong, Yongshuo
Yu, Tingyang
Chavhan, Ruchika
Zhao, Bingchen
Hospedales, Timothy
contents Large language and vision-language models are rapidly being deployed in practice thanks to their impressive capabilities in instruction following, in-context learning, and so on. This raises an urgent need to carefully analyse their robustness so that stakeholders can understand if and when such models are trustworthy enough to be relied upon in any given application. In this paper, we highlight a specific vulnerability in popular models, namely permutation sensitivity in multiple-choice question answering (MCQA). Specifically, we show empirically that popular models are vulnerable to adversarial permutation in answer sets for multiple-choice prompting, which is surprising as models should ideally be as invariant to prompt permutation as humans are. These vulnerabilities persist across various model sizes, and exist in very recent language and vision-language models. Code is available at https://github.com/ys-zong/FoolyourVLLMs.
format Preprint
id arxiv_https___arxiv_org_abs_2310_01651
institution arXiv
publishDate 2023
record_format arxiv
spellingShingle Fool Your (Vision and) Language Model With Embarrassingly Simple Permutations
Zong, Yongshuo
Yu, Tingyang
Chavhan, Ruchika
Zhao, Bingchen
Hospedales, Timothy
Machine Learning
Large language and vision-language models are rapidly being deployed in practice thanks to their impressive capabilities in instruction following, in-context learning, and so on. This raises an urgent need to carefully analyse their robustness so that stakeholders can understand if and when such models are trustworthy enough to be relied upon in any given application. In this paper, we highlight a specific vulnerability in popular models, namely permutation sensitivity in multiple-choice question answering (MCQA). Specifically, we show empirically that popular models are vulnerable to adversarial permutation in answer sets for multiple-choice prompting, which is surprising as models should ideally be as invariant to prompt permutation as humans are. These vulnerabilities persist across various model sizes, and exist in very recent language and vision-language models. Code is available at https://github.com/ys-zong/FoolyourVLLMs.
title Fool Your (Vision and) Language Model With Embarrassingly Simple Permutations
topic Machine Learning
url https://arxiv.org/abs/2310.01651