Prompt Tuning

Prompt tuning learns continuous soft prompt embeddings prepended to inputs while all model weights stay frozen.

Unlike discrete prompts written as words, soft prompts live only in embedding space. They are learned vectors, not tokens you can read. Training moves those vectors until the frozen model produces the desired downstream behavior. 1% of the model, typically 20-100 learned tokens. At inference the soft prompt is prepended and the frozen model runs as usual.

The gap to full fine-tuning shrinks as models get larger. Very large models can reach near-parity. Google's 2021 paper showed an 11B parameter model with prompt tuning matching full fine-tuning on SuperGLUE. The learned prompts are not interpretable as words. They still pack task-specific information that steers the model.

Prompt tuning works when the task is well defined and the target behavior is consistent. It struggles on complex, multi-faceted tasks that need weight-level changes.

Twenty to 100 vectors is a small matrix compared with billions of weights. Gradients update only those vectors. The frozen 11B model in the 2021 SuperGLUE result did not change its attention or feedforward weights. Soft prompts cannot be printed as English, so you cannot copy them into a chat box. They still work when the task has a stable input-output map.

If the task needs new factual circuits or multi-skill routing, weight-level methods such as LoRA tend to win. Prompt tuning learns soft prompt vectors. At large scale it can match full fine-tuning with far fewer trained parameters.