Can soft prompts be trained without changing the base model?

Yes. That is the central arrangement in the input-level method described here.

How does prefix tuning differ?

Related prefix methods can condition internal layers as well as the input. The location and parameterization of the added values distinguish implementations.

Can a soft prompt be reused for another task?

Transfer is possible in some settings, but useful performance must be tested. Compatibility of dimensions alone does not establish task suitability.

AI Glossary / Prompting and human-AI interaction / Advanced

Soft prompt tuning

Definition

Soft prompt tuning adapts a pretrained model by learning a small set of continuous prompt vectors while keeping the base model's weights fixed. The vectors condition the model's responses to a task. Unlike ordinary text prompts, they are numerical parameters that need not correspond to readable vocabulary tokens.

What it means

The word soft distinguishes trainable continuous representations from hard prompts made of chosen words. Instead of searching only among possible sentences, an optimizer can adjust each coordinate of the prompt vectors to reduce a training loss.

In the common input-level formulation, the learned vectors precede the embeddings of the actual input. The same frozen model can then be paired with different soft prompts for different tasks.

Soft prompt tuning is often simply called prompt tuning. Broader terminology sometimes groups related continuous-prefix methods together, so a description should state where the learned values enter the model and which parameters are updated. This entry focuses on input-level soft prompts.

How it works

Prepare task examples and target outputs, select a prompt length and initialize its vectors. The input embeddings and soft prompt pass through the frozen model. A loss measures how well the generated predictions match the targets.

Gradients flow through the computation to the prompt vectors, and an optimizer changes those vectors. The base weights remain fixed, but their numerical values still determine how the prompt affects predictions. Disabling every gradient through the model would prevent ordinary gradient-based prompt learning.

After training, evaluation uses examples not used to select the prompt. Serving requires the compatible base model and the learned prompt together. A stored soft prompt is not a standalone language model, and inserting its numerical values into an ordinary chat message is not equivalent to supplying them through an embedding interface.

Three practical examples

A classifier learns a task-specific soft prompt that steers a shared language model toward the desired label vocabulary.
A document application maintains separate learned prompts for different extraction tasks while reusing a compatible frozen backbone.
A researcher compares random initialization with initialization from vocabulary embeddings, keeping the training data and evaluation protocol fixed.

Think of it this way

Think of adjusting the input controls of a fixed machine. The machine's internal parts remain in place, but different settings can change its response to the same material.

The settings do not need to have readable names. Likewise, a soft prompt can influence behavior without being a hidden English instruction that can be recovered by simply looking up the nearest words.

Worked example

Consider an original scalar teaching model that illustrates gradient flow, not a transformer implementation. Let the fixed input be x = 0.5 and the fixed model coefficient be w = 2. Let p be the single trainable prompt-like value. Define the logit z = w(x + p), probability q = 1/(1 + exp(-z)), and target y = 1.

Initially p = 0, so z = 1 and q is approximately 0.73106. The negative log-likelihood loss is -ln(q), approximately 0.31326.

For this model, the derivative of the loss with respect to p is w(q - 1), approximately -0.53788. Using a learning rate of 0.1 gives p_new = 0 - 0.1(-0.53788), approximately 0.05379.

The coefficient w is still 2. The new logit is approximately 1.10758, the new probability is approximately 0.75168, and the loss falls to approximately 0.28545. The output changed because the prompt-like parameter changed, even though the model coefficient did not.

Real soft prompts contain many vector coordinates and interact with a much more complex model. This simple calculation demonstrates the distinction between freezing model parameters and blocking all gradient flow. A reduction on one training example does not establish useful generalization.

Illustrative case study

Fictional case: Birch Archives evaluates soft prompts for classifying internal records. Its team freezes a base model, trains a prompt on reviewed examples and reserves separate development and test collections.

The engineers inspect which parameters the optimizer updates, rather than assuming a configuration labeled "frozen" is correct. They save the prompt alongside the base-model identifier and label mapping, and test whether changing prompt length affects accuracy and serving cost.

When a newer backbone becomes available, they treat prompt compatibility as an empirical question. This hypothetical case illustrates implementation checks and claims no measured storage or accuracy advantage.

Common misconceptions and mistakes

Soft means an informal or optional instruction: it refers to continuous learned representations.
A soft prompt is just a long text prompt: its vectors need not match vocabulary embeddings.
Freezing the base model eliminates all training computation: the model still participates in the forward and gradient calculations.
A readable nearest-word sequence is equivalent to the learned vectors: replacing vectors with discrete tokens changes the input.
A small trainable component cannot overfit: limited or unrepresentative task data can still produce poor generalization.

Business use

Soft prompt tuning may help when task-specific parameter storage is important and a suitable shared model is available. It requires a training and serving interface that can use continuous prompts; ordinary access to a text-generation endpoint may be insufficient.

Compare it with textual prompting, full fine-tuning and other parameter-efficient methods on the actual task. Measure total memory, latency and quality rather than only the prompt's parameter count. Keep version information and evaluation evidence with each prompt, since a small file can still be a critical part of the deployed model configuration.

Check your understanding

The base weights are frozen, but gradients are computed through their operations to update the prompt. Is that contradictory?

Answer: No. A parameter can influence a computation without being updated. Gradients with respect to the prompt still depend on the frozen model's operations.

Related terms

Prompt tuning
Embedding
Backpropagation
Overfitting
soft prompt
frozen backbone
prefix tuning
parameter-efficient adaptation
initialisation

Learn next

Inspect a small implementation and identify exactly which tensors are trainable. Check their dimensions and verify that base weights remain unchanged after an update. Then evaluate on held-out inputs and distinguish training-loss improvement from task-level generalization.

Training loss

Sources

Lester, Al-Rfou and Constant, The Power of Scale for Parameter-Efficient Prompt Tuning, EMNLP 2021, Section 2: https://aclanthology.org/2021.emnlp-main.243.pdf

Original research implementation, retained as an archived repository: https://github.com/google-research/prompt-tuning

The scalar gradient calculation and Birch Archives case are original hypothetical illustrations. The scalar model is a teaching analogy, not the architecture used in the cited experiments.

Previous: Soft margin Next: Softmax