Soft Prompt
A soft prompt is a learned vector (or short sequence of vectors) in embedding space, prepended to the model input. Unlike a hard prompt, it does not map to vocabulary tokens you can read as words.
Soft prompts are the parameters updated in prompt tuning: the base model stays frozen; only the prompt embeddings are trained. They often beat hand-written hard prompts on a target task, at the cost of interpretability—you cannot copy-paste the soft prompt into a chat UI as text.
Trade-offs
- Tiny trainable footprint vs full fine-tuning or even LoRA
- Task-specific; usually not human-editable
- Must be stored and loaded with the serving path that knows how to inject them
Related: PEFT methods that adapt a model without full fine-tuning. Contrast with prompt engineering, which edits readable hard prompts only.