A family of ways to adapt a large model to a new task by training only a small number of extra or selected settings, leaving the rest untouched. Retraining a whole model is expensive and produces a full-sized copy for every task. These methods cost far less to run and store, so one base model can carry many small add-ons for different jobs.
Technical view
Covers adapters (small bottleneck layers inserted into each block), prefix and prompt tuning (learned vectors prepended to the input or to each layer’s keys and values), and low-rank methods such as Low-Rank Adaptation (‘LoRA’), which learn a low-rank update to frozen weight matrices. Typically well under a few per cent of parameters are trained, cutting optimiser memory and allowing many task modules to be swapped at serving time. Combined with quantisation it makes fine-tuning feasible on a single graphics card. Trade-offs are a possible quality gap against full fine-tuning on harder shifts, and, on some evidence, less capacity than full fine-tuning to absorb substantially new knowledge.
Related
Part of AI Terminology Index, section: Training and Post-training.