BCMT Architecture Decouples Local Attention From Global Context for Long-Context Modeling

The Blockwise Causal Memory Transformer (BCMT) replaces quadratic dense self-attention with local block attention and an adaptive exponential causal memory. This approach enables efficient long-range context propagation while remaining fully parallelizable and compatible with standard implementations.