2026-04-09 CAD: Disaggregating Core Attention for Efficient Long-context Language Model Training