RLPrompt

Overview

RLPrompt is a discrete prompt-optimization method that trains a lightweight policy network to produce prompt tokens for a frozen language model. It is the clean canonical paper for reinforcement learning over prompt text rather than over model weights.

Why it matters

It matters because it makes the basic move explicit: the prompt is a learnable external artifact. For harness design, that is a decisive conceptual shift even when the optimized object is still only one prompt.

Distinctive trait

Its distinctive trait is that the learned prompts are often ungrammatical but still effective and somewhat transferable across models. The optimizer is clearly doing something, but not necessarily something a human would enjoy reading.

Relationships

Read RLPrompt with tempera, autodspy, dspy, gepa, prompt-optimization-and-dspy-follow-ups, and prompt-optimization-timeline-and-harness-lessons. It is also a useful contrast class for reflexion, where the learned artifact is memory rather than a single prompt.