As large language models (LLMs) process increasingly long sequences, attention becomes a major computational bottleneck. Sparse attention mostly targets infe…
机构：Microsoft
来源：arXiv 2610.31911 | AI4Papers 论文推荐平台