Mixture-of-Experts (MoE) models expand model capacity without a proportional increase in training compute, but increasing sparsity makes reliable hyperparame…
机构:蚂蚁
来源:arXiv 2609.08690 | AI4Papers 论文推荐平台