One Token at a Time Multimodal Attention Analysis

The OTaT study tracks how multimodal large language models shift attention between image, text, and instruction tokens during autoregressive generation. The research identifies consistent attention-switching patterns across different model sizes and architectures during single-response generation.