Training-Free Dynamic Upcycling of Expert Language Models
arXiv:2603.29765v1 Announce Type: cross
Abstract: Large Language Models (LLMs) have achieved remarkable performance on a wide range of specialized tasks, exhibiting strong problem-solving capabilities. However, training these models is prohibitively e…