Abstract page for arXiv paper 2608.20427: BF1: A Causal Dyadic Sparse-Attention Retrofit for Efficient Long-Context Transformers