Decoder. plain-English AI glossary

Soft prompts

▲ Rising

Embedding vectors that act like prompts — learned during training rather than written as text.

Think of it like

A melody hummed (learned) vs. a song title spoken (written).

Example

Instead of "Classify this as positive or negative," the model processes learned vectors that implicitly instruct classification.

How it actually works

Soft prompts are prefixed to the model's input embeddings. They're optimized via backprop. Advantage: they exploit the model's embedding space fully. Disadvantage: black box — you can't inspect or transfer them. Effective for task-specific adaptation without fine-tuning the full model.

For product teams

Lightweight adaptation for domain-specific tasks.

For engineers

Implement as a learnable embedding layer. Train on task-specific data with a loss function. Validate generalization.

Prompt tuning — Core technique in prompt tuning.
Prefix tuning — Related to prefix tuning.

Read anything AI without the jargon

Look up any term in plain English, or save terms as you read with the free Chrome extension.

Open Decoder Add to Chrome