Unlocking Causal Attention into Modality-Mutual Attention for Multimodal LLMs (github.com)

AKI, a novel MLLM that unlocks causal attention into modality-mutual attention (MMA) to enable image tokens to attend to text tokens.