Cambrian-1

New method measures encoder roles in multi-encoder vision-language models

Researchers have developed a new method to analyze the roles of different encoders in multi-encoder large vision-language models (LVLMs). By retraining subsets of five common vision encoders on the Cambrian-1 benchmark,…