Researchers have introduced Simple Attention Sparsification (SAS), a novel post-training method designed to mitigate the quadratic computational complexity of Transformer models by selecting a subset of context units. Unlike existing trainable sparsification techniques that rely on lightweight selectors and hard Top-K selection—which often leads to gradient blockage and misalignment between dense attention weights and predictive utility—SAS optimizes context ranking end-to-end using the standard language modeling loss. The method achieves this by injecting continuous selector scores directly into the attention logits during training, allowing for seamless backpropagation. To ensure stability and performance, the authors implement three critical design choices: integrating the gate within the log-space of the attention softmax, utilizing normalized softmax gates to calibrate historical context against the current block, and maintaining continuous scores to capture relative priorities. To facilitate efficient long-sequence training, the authors developed a memory-efficient Triton kernel that integrates SAS into FlashAttention-style computations. Experimental results across reasoning, long-context understanding, and agentic tasks demonstrate that SAS consistently outperforms current trainable sparse attention baselines. The performance gains are particularly significant under constrained attention budgets, proving that SAS more effectively ranks the most impactful context units for downstream tasks compared to methods that merely distill layer-wise dense distributions.