Cohere Labs - Ashmal Vayani, MSc student, Univ. of Central Florida

For all the promise of “global” multimodal AI, most systems still speak to a narrow slice of the world. This talk traces a path to genuinely multilingual, culturally grounded models through three steps: first, ALM-Bench reframes evaluation by testing 100 languages across culturally situated tasks, revealing where today’s LMMs falter, especially on low-resource scripts. Next, ViMUL moves beyond images to video, pairing a diverse 14-language, 15-domain benchmark with a balanced baseline to show how training and evaluation can align for robust multilingual video understanding. Finally, we examine language itself as a causal factor: a cross-lingual T2I study where grammatical gender shifts visual outputs, surfacing a new axis of bias. Together, these pieces offer a story and a blueprint for inclusive, reliable multimodal systems.

I am an MSc. student in the College of Engineering and Computer Science department at the University of Central Florida. I am a member of the Center for Research in Computer Vision (CRCV) Lab advised by Prof. Mubarak Shah.

Previously, I was a Research Engineer in the Computer Vision Department, affiliated with the IVAL-Lab at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI).

Add event to calendar

Apple Google Office 365 Outlook Outlook.com Yahoo