Soft Prompt

Definitions

1. A learnable prompt representation, typically implemented as trainable embeddings, that influences model behavior without changing the full model weights. Soft prompts are a parameter-efficient way to adapt models to tasks.

   In plain English: A trainable embedded prompt used to steer model behavior without changing the whole model.

   Example: "They used a soft prompt to specialize the model for biomedical extraction while keeping the base weights frozen."

Related Terms

• system prompt
• Prompt Leaking
• Chain of Density
• Chain of Verification
• Context Stuffing
• Few-Shot
• Low-Rank Adaptation
• Unlocking
• Zero-Shot