Skip to main content
All terms
Optimization

Adapter

Small trainable modules inserted into a frozen model so it can learn new tasks cheaply.

Definition

Adapters are small neural network modules — a tiny stack that shrinks the data, transforms it, then expands it back — inserted between the layers of a frozen (unchanged) pretrained model. Only the adapter parameters are trained, leaving the base weights untouched. This parameter-efficient fine-tuning approach lets one backbone serve many tasks by swapping adapters, cutting storage and compute compared with full fine-tuning. LoRA and IA3 are modern variants of the idea.