parameter-efficient fine-tuning (PEFT)

Full fine-tuning updates every weight, which means a separate copy of a giant model per task, a large optimizer state, and gradients for billions of parameters. Parameter-efficient fine-tuning is the umbrella name for methods that adapt a model by training only a tiny fraction of its parameters — or a small set of new ones bolted on — while the vast pretrained backbone stays frozen. The premise is that task adaptation is intrinsically low-dimensional, so you should not have to pay full price to capture it.

The family splits into a few mechanisms. Additive methods insert new trainable modules and leave originals untouched: adapter layers, prefix tuning, and soft prompts. Reparameterization methods express the weight update in a compact form, most famously the low-rank factorization of LoRA and its descendants DoRA and QLoRA. Selective methods unfreeze only a chosen sliver of existing parameters, such as just the biases. All share the same accounting win: trainable parameters and optimizer memory shrink by one to three orders of magnitude.

The practical consequences are large. Many tasks can share one frozen base with swappable lightweight modules, cutting storage and enabling multi-tenant serving; cheaper hardware suffices; and forgetting is milder because the base is preserved. The tradeoff is a usually-small accuracy gap versus full fine-tuning, and the fact that no single PEFT method dominates — the right choice depends on task, scale, and serving constraints.