IISAN-Versa

A decoupled parameter-efficient adaptation framework for symmetric and asymmetric multimodal foundation models in recommendation.

IISAN-Versa adapts multimodal foundation models for sequential recommendation while reducing the memory and training costs associated with full fine-tuning.

Its decoupled side-adapter design supports both symmetric and asymmetric text-image encoder combinations. The project also covers alignment between encoders with different depths and representation dimensions.