CASA: Cross-Attention via Self-Attention for Efficient Vision-Language Fusion

ARC-Encoder: learning compressed text representations for large language models