Sub-Quadratic Block-Sparse Attention via Odd Binomial Fractal Tree Topologies

Standard full self-attention in Transformer architectures incurs a quadratic computational and memory complexity, O(N^2), rendering long-context scaling computationally prohibitive. While heuristic and randomized sparse attention patterns alleviate computational overhead, they either suffer from severe context blind spots or introduce stochastic, non-deterministic gradient flows that fail to preserve global associative recall. In this work, we introduce Fractal Block-Sparse Attention, a deterministic, hardware-aligned sparse attention mechanism governed by odd binomial fractal tree dynamics...