Learned sparse attention, used in production models such as DeepSeek-V4 and GLM-5, lets each query attend to only k keys chosen by a lightweight indexer, but…
机构:NVIDIA
来源:arXiv 2611.02508 | AI4Papers 论文推荐平台