Prompt Tuning: Definition and Examples

Prompt tuning is a method for adapting large language models (LLMs) that stands out from classical fine-tuning due to its efficiency and lightweight nature. Instead of modifying all billions of parameters of a model, only a small vector of virtual tokens — called "soft prompts" — is trained and prepended to the model's input. These tokens do not correspond to actual words in the vocabulary but are numerical representations optimized via backpropagation for a specific task.

This approach was popularized by the Google research paper "The Power of Scale for Parameter-Efficient Prompt Tuning" (Lester et al., 2021), which demonstrated that for sufficiently large models, prompt tuning achieves performance comparable to full fine-tuning while modifying only a tiny fraction of the parameters (often less than 0.1%). This makes the technique particularly interesting for production deployments where a single base model must serve multiple different tasks.

Concretely, prompt tuning works in three steps: initialize a set of embedding vectors (the soft prompts), prepend them to each training example, then optimize only these vectors via gradient descent while the model remains frozen. The result is a file of a few kilobytes that encodes the learned "instruction" for the target task.

It is important not to confuse prompt tuning with prompt engineering. Prompt engineering involves manually crafting natural language instructions, while prompt tuning uses machine learning to discover optimal representations that are not interpretable by humans. The two approaches are complementary: prompt engineering is accessible to everyone, while prompt tuning requires a training pipeline but offers measurable performance gains on repetitive tasks.