Keumseo Ryum

CT-Merging: Consensus Directions and Task-Level Scaling for LoRA Adapter Merging

Keumseo Ryum, Joonhyuk Kang

LoRA adapters provide an efficient way to specialize a pretrained model for many downstream tasks, but deploying one adapter per task requires adapter storage and task selection at inference time. Model merging addresses this issue by combining independently trained adapters into…Read more