Definition
Prompt tuning adapts a pretrained model by learning a small sequence of continuous vectors, often called a soft prompt. During training, these vectors are placed alongside the input token embeddings. Gradients update the soft prompt while the base model’s weights stay frozen. At inference time, the same learned vectors accompany each input for that task.
Unlike a prompt written in ordinary language, the learned vectors do not have to correspond to readable words. This makes prompt tuning a parameter-efficient model adaptation method, even though the adaptation enters through the model’s input.
Simple example
A team wants a model to classify support tickets as network, storage, or identity. It trains a soft prompt on reviewed tickets and their correct labels, leaving the model itself frozen. For each new ticket, the serving system supplies the learned vectors with the ticket text. It can use a different soft prompt for incident summaries while reusing the same base model.
Why it matters
Only the prompt vectors are trainable. Any per-parameter optimizer state is limited to those vectors. At inference time, each task needs only its learned prompt parameters alongside the base model. That can make training and storing several adaptations cheaper than updating and saving a full model for each task. Backpropagation still passes through the frozen model, so the training work does not disappear. The serving system must also apply the right soft prompt to each request.
One important nuance
Prompt tuning is different from editing a text prompt or adding few-shot examples. Those approaches supply readable tokens at request time and need no training run. A soft prompt is learned from task data and cannot be inspected as a set of plain-language instructions. LoRA instead trains low-rank factors that produce updates to selected weight matrices. The pretrained weights stay frozen.
Soft prompts also require serving support for supplying embeddings. A text-only model API cannot accept one as a string. Freezing the base weights does not guarantee that a soft prompt will match full fine-tuning on a given task. Compare quality on representative inputs, and version each soft prompt with the base model used to train it.