Designing expressive sequence layers with efficient inference remains a central challenge in modern machine learning. Standard softmax attention achieves exc…
机构:Stanford
来源:arXiv 2609.39034 | AI4Papers 论文推荐平台