Definition
A soft prompt is a set of learnable vectors, often called virtual tokens, that are concatenated with a model's input embeddings and trained to steer a frozen language model toward a task. It is the continuous-valued counterpart to a hard prompt, the ordinary handwritten text instruction you type into a model.
Hard versus soft prompts
Hard prompts are discrete: they are made of real tokens from the vocabulary and are human-readable, but crafting a good one is laborious and often brittle. Soft prompts replace that manual effort with optimization. Because the virtual tokens live in continuous embedding space rather than mapping to actual words, they are not human-readable, but they can be tuned by gradient descent to outperform handcrafted prompts. Crucially, only the soft-prompt tensor is updated during training; the model's own weights stay frozen, so one base model can serve many tasks, each with its own small learned prompt.
The soft-prompt family
Several parameter-efficient fine-tuning methods are built on this idea and differ mainly in where the learnable vectors are placed and how they are produced. Prompt tuning adds them only at the input embedding layer; prefix tuning injects them into every transformer layer via a reparameterization network; and P-tuning generates them through a small prompt encoder and can insert them anywhere in the sequence.
For sovereign builders fine-tuning open-weight models locally, soft prompts are one of the cheapest ways to specialize a model, since the trained artifact is often just a few kilobytes of vectors rather than a full model copy.
In Simple Terms
A soft prompt is a set of learnable vectors, often called virtual tokens, that are concatenated with a model’s input embeddings and trained to steer…