Sensitivity of Generative VLMs to Semantically and Lexically Altered Prompts

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Dumpala, Sri Harsha, Jaiswal, Aman, Sastry, Chandramouli, Milios, Evangelos, Oore, Sageev, Sajjad, Hassan
Natura: Preprint
Pubblicazione: 2024
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866929547448942592
author Dumpala, Sri Harsha
Jaiswal, Aman
Sastry, Chandramouli
Milios, Evangelos
Oore, Sageev
Sajjad, Hassan
author_facet Dumpala, Sri Harsha
Jaiswal, Aman
Sastry, Chandramouli
Milios, Evangelos
Oore, Sageev
Sajjad, Hassan
contents Despite the significant influx of prompt-tuning techniques for generative vision-language models (VLMs), it remains unclear how sensitive these models are to lexical and semantic alterations in prompts. In this paper, we evaluate the ability of generative VLMs to understand lexical and semantic changes in text using the SugarCrepe++ dataset. We analyze the sensitivity of VLMs to lexical alterations in prompts without corresponding semantic changes. Our findings demonstrate that generative VLMs are highly sensitive to such alterations. Additionally, we show that this vulnerability affects the performance of techniques aimed at achieving consistency in their outputs.
format Preprint
id arxiv_https___arxiv_org_abs_2410_13030
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Sensitivity of Generative VLMs to Semantically and Lexically Altered Prompts
Dumpala, Sri Harsha
Jaiswal, Aman
Sastry, Chandramouli
Milios, Evangelos
Oore, Sageev
Sajjad, Hassan
Computer Vision and Pattern Recognition
Computation and Language
Machine Learning
Despite the significant influx of prompt-tuning techniques for generative vision-language models (VLMs), it remains unclear how sensitive these models are to lexical and semantic alterations in prompts. In this paper, we evaluate the ability of generative VLMs to understand lexical and semantic changes in text using the SugarCrepe++ dataset. We analyze the sensitivity of VLMs to lexical alterations in prompts without corresponding semantic changes. Our findings demonstrate that generative VLMs are highly sensitive to such alterations. Additionally, we show that this vulnerability affects the performance of techniques aimed at achieving consistency in their outputs.
title Sensitivity of Generative VLMs to Semantically and Lexically Altered Prompts
topic Computer Vision and Pattern Recognition
Computation and Language
Machine Learning
url https://arxiv.org/abs/2410.13030