_version_ 1866910921456091136
author Salazar, Israfel
Burda, Manuel Fernández
Islam, Shayekh Bin
Moakhar, Arshia Soltani
Singh, Shivalika
Farestam, Fabian
Romanou, Angelika
Boiko, Danylo
Khullar, Dipika
Zhang, Mike
Krzemiński, Dominik
Novikova, Jekaterina
Shimabucoro, Luísa
Imperial, Joseph Marvin
Maheshwary, Rishabh
Duwal, Sharad
Amayuelas, Alfonso
Rajwal, Swati
Purbey, Jebish
Ruby, Ahmed
Popovič, Nicholas
Suppa, Marek
Wasi, Azmine Toushik
Kadiyala, Ram Mohan Rao
Tsymboi, Olga
Kostritsya, Maksim
Moakhar, Bardia Soltani
Merlin, Gabriel da Costa
Coletti, Otávio Ferracioli
Shiviari, Maral Jabbari
fard, MohammadAmin farahani
Fernandez, Silvia
Grandury, María
Abulkhanov, Dmitry
Sharma, Drishti
De Mitri, Andre Guarnier
Marchezi, Leticia Bossatto
Heydari, Setayesh
Obando-Ceron, Johan
Kohut, Nazar
Ermis, Beyza
Elliott, Desmond
Ferrante, Enzo
Hooker, Sara
Fadaee, Marzieh
author_facet Salazar, Israfel
Burda, Manuel Fernández
Islam, Shayekh Bin
Moakhar, Arshia Soltani
Singh, Shivalika
Farestam, Fabian
Romanou, Angelika
Boiko, Danylo
Khullar, Dipika
Zhang, Mike
Krzemiński, Dominik
Novikova, Jekaterina
Shimabucoro, Luísa
Imperial, Joseph Marvin
Maheshwary, Rishabh
Duwal, Sharad
Amayuelas, Alfonso
Rajwal, Swati
Purbey, Jebish
Ruby, Ahmed
Popovič, Nicholas
Suppa, Marek
Wasi, Azmine Toushik
Kadiyala, Ram Mohan Rao
Tsymboi, Olga
Kostritsya, Maksim
Moakhar, Bardia Soltani
Merlin, Gabriel da Costa
Coletti, Otávio Ferracioli
Shiviari, Maral Jabbari
fard, MohammadAmin farahani
Fernandez, Silvia
Grandury, María
Abulkhanov, Dmitry
Sharma, Drishti
De Mitri, Andre Guarnier
Marchezi, Leticia Bossatto
Heydari, Setayesh
Obando-Ceron, Johan
Kohut, Nazar
Ermis, Beyza
Elliott, Desmond
Ferrante, Enzo
Hooker, Sara
Fadaee, Marzieh
contents The evaluation of vision-language models (VLMs) has mainly relied on English-language benchmarks, leaving significant gaps in both multilingual and multicultural coverage. While multilingual benchmarks have expanded, both in size and languages, many rely on translations of English datasets, failing to capture cultural nuances. In this work, we propose Kaleidoscope, as the most comprehensive exam benchmark to date for the multilingual evaluation of vision-language models. Kaleidoscope is a large-scale, in-language multimodal benchmark designed to evaluate VLMs across diverse languages and visual inputs. Kaleidoscope covers 18 languages and 14 different subjects, amounting to a total of 20,911 multiple-choice questions. Built through an open science collaboration with a diverse group of researchers worldwide, Kaleidoscope ensures linguistic and cultural authenticity. We evaluate top-performing multilingual vision-language models and find that they perform poorly on low-resource languages and in complex multimodal scenarios. Our results highlight the need for progress on culturally inclusive multimodal evaluation frameworks.
format Preprint
id arxiv_https___arxiv_org_abs_2504_07072
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Kaleidoscope: In-language Exams for Massively Multilingual Vision Evaluation
Salazar, Israfel
Burda, Manuel Fernández
Islam, Shayekh Bin
Moakhar, Arshia Soltani
Singh, Shivalika
Farestam, Fabian
Romanou, Angelika
Boiko, Danylo
Khullar, Dipika
Zhang, Mike
Krzemiński, Dominik
Novikova, Jekaterina
Shimabucoro, Luísa
Imperial, Joseph Marvin
Maheshwary, Rishabh
Duwal, Sharad
Amayuelas, Alfonso
Rajwal, Swati
Purbey, Jebish
Ruby, Ahmed
Popovič, Nicholas
Suppa, Marek
Wasi, Azmine Toushik
Kadiyala, Ram Mohan Rao
Tsymboi, Olga
Kostritsya, Maksim
Moakhar, Bardia Soltani
Merlin, Gabriel da Costa
Coletti, Otávio Ferracioli
Shiviari, Maral Jabbari
fard, MohammadAmin farahani
Fernandez, Silvia
Grandury, María
Abulkhanov, Dmitry
Sharma, Drishti
De Mitri, Andre Guarnier
Marchezi, Leticia Bossatto
Heydari, Setayesh
Obando-Ceron, Johan
Kohut, Nazar
Ermis, Beyza
Elliott, Desmond
Ferrante, Enzo
Hooker, Sara
Fadaee, Marzieh
Computation and Language
Computer Vision and Pattern Recognition
The evaluation of vision-language models (VLMs) has mainly relied on English-language benchmarks, leaving significant gaps in both multilingual and multicultural coverage. While multilingual benchmarks have expanded, both in size and languages, many rely on translations of English datasets, failing to capture cultural nuances. In this work, we propose Kaleidoscope, as the most comprehensive exam benchmark to date for the multilingual evaluation of vision-language models. Kaleidoscope is a large-scale, in-language multimodal benchmark designed to evaluate VLMs across diverse languages and visual inputs. Kaleidoscope covers 18 languages and 14 different subjects, amounting to a total of 20,911 multiple-choice questions. Built through an open science collaboration with a diverse group of researchers worldwide, Kaleidoscope ensures linguistic and cultural authenticity. We evaluate top-performing multilingual vision-language models and find that they perform poorly on low-resource languages and in complex multimodal scenarios. Our results highlight the need for progress on culturally inclusive multimodal evaluation frameworks.
title Kaleidoscope: In-language Exams for Massively Multilingual Vision Evaluation
topic Computation and Language
Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2504.07072