Scalable AI Lab (SAIL) · Soongsil University · PI: Prof. Hyesong Choi

Efficient and Trustworthy Multimodal Foundation Models

From Visual Perception to Physical AI

We build vision-language and multimodal foundation models that are efficient to train, reliable to deploy, and ready to move from the screen into the physical world. Our core idea is Corruption-to-Reconstruction (C2R): corrupting what a model sees (masking, noise, embedding perturbation) and learning to restore it yields stronger representations with less compute and less data.