soft prompt tuning

Normally you steer a model with words you type — a hard prompt. Soft prompt tuning learns the steering instead. It prepends a handful of trainable vectors to your input — vectors that live in the model's embedding space but correspond to no actual words — and tunes only those, leaving the entire model frozen.

These continuous virtual tokens are optimised by gradient descent so that, once prepended to a task input, they coax the right behaviour out of the frozen model. Because you train only a few thousand numbers, it is extremely lightweight, and a single base model can host many learned prompts at once. A related variant, prefix tuning, injects such learned vectors at every layer rather than only at the input.

Soft prompts are powerful but can be finicky to train, and they are less interpretable than text prompts — you cannot read off what they say.