Prompt tuning

Two different things share the word “prompt”

Every earlier lesson in this roadmap has been about writing better natural-language text: choosing words, giving examples, structuring instructions. All of that is prompt engineering — a human skill applied to a frozen model, with no training involved. Prompt tuning is a different thing entirely, despite the name. It is a parameter-efficient fine-tuning (PEFT) technique: a way to adapt a large language model to a specific task by training a small number of new parameters, while leaving the model’s own weights untouched.

The confusion is understandable, because the two produce something that looks similar at inference time — a sequence prepended to your real input, steering the model’s output. But the mechanism and the cost structure are nothing alike. One is something you type. The other is something you train.

What actually gets trained

A soft prompt is a short sequence of continuous embedding vectors — typically 10 to 100 of them — prepended to the input’s token embeddings before they enter the model. These vectors do not correspond to any real token in the model’s vocabulary; you cannot decode them back into words, because they were never words to begin with. They start as random (or heuristically initialized) numbers and are updated by ordinary gradient descent, exactly the way a model’s own weights are updated during training, except here every weight in the base model is frozen and only the soft-prompt embeddings receive gradients.

Training requires a labeled dataset for the target task — input examples paired with the outputs you want — and a loss function comparing the model’s predictions against those labels, backpropagated only into the soft prompt. After enough steps, the learned embeddings reliably push the frozen model toward the desired behavior for that specific task. A new task means training a new soft prompt from scratch; the technique produces one small, task-specific artifact per task rather than one adapted model.

Why bother, if you already have prompt engineering

Manual prompting has a ceiling. Natural language is a surprisingly lossy way to specify exact, repeatable behavior — a human-written instruction can always be phrased a different way, and models drift in how literally they follow phrasing across versions. Prompt tuning searches a much larger, continuous space than any human would try by hand, and for a narrow, well-defined task it can reliably outperform the best discrete prompt anyone writes, because gradient descent, not trial and error, is doing the search. It also decouples the task adaptation from the prompt text shown to an end user, if any — the soft prompt lives behind the scenes.

The cost is what it has always been with any trained artifact: you need labeled data, a training loop, and infrastructure to run it, none of which a hand-written prompt requires. For a one-off task, or one without a labeled dataset, prompt engineering is still the right first move — it costs nothing but iteration time.

Where it sits on the spectrum

Prompt tuning is one point on a spectrum of techniques that trade an increasing amount of compute and data for an increasing amount of control over model behavior:

• Prompt engineering — no training, discrete text, cheapest and fastest to iterate, bounded by what language alone can specify.
• Prompt tuning — trains a small set of continuous embeddings per task, frozen base model, needs labeled data and a training loop.
• Prefix-tuning and p-tuning — close relatives that train small continuous vectors injected at every transformer layer (not just the input) for more expressive control, at similar cost.
• LoRA — trains small low-rank update matrices inside the model’s own weight matrices rather than prepending anything, still with the base weights frozen.
• Full fine-tuning — updates every weight in the model, the most capable and most expensive option, covered in the next lesson.

None of these are covered in depth here — the point is to place prompt tuning correctly: well past plain prompting in terms of what it requires, well short of full fine-tuning in terms of what it costs. If a task has a clear, stable definition and a labeled dataset but doesn’t justify retraining an entire model, prompt tuning is usually the first PEFT technique worth trying before reaching for something heavier.