SAIL: Sparse Autoencoders for Interpretable Alignment of Human Vision and Multimodal Large Language Models
Tan Gao, Luyuan Zhang, Yifei Wang, Yunjia Wang, Haotian Zhou, Mufan Xue, Guoyuan Yang
SAIL aligns human vision and MLLM representations using sparse autoencoders, revealing semantic alignment along the ventral visual pathway and action-related semantics in specific brain regions.
bioRxiv · Oct 6, 2026 · PreprintNeuroscience