[Paper] Delta-LLaVA: Base-then-Specialize Alignment for Token-Efficient Vision-Language Models
Multimodal Large Language Models (MLLMs) combine visual and textual representations to enable rich reasoning capabilities. However, the high computational cost ...