What is KOSMOS-2?

Опубликовано: 11 Март 2026
на канале: Samuel Albanie

4,738

162

"KOSMOS-2: Grounding Multimodal Large Language Models to the World" is a new preprint from Microsoft research that illustrates multimodal grounding abilities in a large vision-language model.

Timestamps:
00:00 - KOSMOS-2
00:12 - Grounding Multimodal Large Language Models to the World
01:25 - KOSMOS-1
02:01 - KOSMOS-2 overview
03:42 - The Grounded Image-Text Pairs (GrIT) dataset
05:07 - Kosmos-2: Model and training details
07:00 - Testing the model on some tricky images
07:57 - Hallucinations (Flamingo reference)
09:34 - Phrase grounding
09:59 - Referring expression comprehension
10:43 - Flamingo - where art thou?
11:56 - Language tasks
12:20 - The Ethics Statement
13:43 - Closing thoughts

Paper link: https://arxiv.org/abs/2306.14824
Demo: https://d20f0c97b1b2ee15.gradio.app/

Topics: #LLMs #ai #microsoft #KOSMOS-2

For related content:
Twitter:   / samuelalbanie
Research lab: https://caml-lab.com/
personal webpage: https://samuelalbanie.com/
YouTube:    / @samuelalbanie1
TikTok:   / samuelalbanie
Instagram:   / samuelalbanie
LinkedIn:   / samuel-albanie
Discord server for filtir:   / discord

(Optional) if you'd like to support the channel:
https://www.buymeacoffee.com/samuelal...
  / samuel_albanie

Acknowledgements:
Image credit for the boat race: https://commons.wikimedia.org/wiki/Fi...