Prompt Tuning

Prompt tuning is a Parameter-Efficient Fine-Tuning (PEFT) technique in which a frozen model is adapted by learning a small set of trainable embeddings that function like a soft prompt. These embeddings are prepended to the input at the embedding level rather than written as literal human-readable text.

The idea is similar to prompting, but instead of hand-writing instructions such as "Summarize this text in bullet points," the system learns a continuous prompt representation that the model responds to more reliably. Because only the prompt embeddings are trained, the number of trainable parameters is tiny.

Advantages

Prompt tuning is one of the lightest-weight adaptation methods available. It is cheap to store, easy to swap, and attractive when you need very small task-specific checkpoints. It can work well for simple steering tasks or for large models that are already highly capable and only need a small nudge.

Disadvantages

Prompt tuning is often less expressive than LoRA or adapter layers, especially for harder tasks requiring deeper behavioral changes. It can also be fragile: performance may depend heavily on model scale, prompt length, and training setup. In smaller or weaker models, prompt tuning may underperform more direct adaptation methods.

Example

A company might learn one virtual prompt that makes a base model answer as a terse support bot and another that makes it respond as a formal legal assistant, without modifying the underlying model weights.