Technique that lets models weight different parts of input sequences, central to transformers and modern LLMs.