TEMPERA

Overview

TEMPERA is a runtime prompt-adaptation method that trains a reinforcement-learning policy to edit an existing prompt per query at test time. Instead of searching once for a single global prompt, it operates over a structured action space spanning instruction phrases, exemplar selection, and verbalizer choice.

Why it matters

It matters because it is a clean anchor for the runtime-adaptation branch of prompt optimization. For harness engineering, the important distinction is that it adapts the live prompt instance without updating model weights or necessarily promoting a durable artifact.

Distinctive trait

Its distinctive trait is interpretable test-time prompt editing over human-seeded prompt components rather than opaque prompt synthesis from scratch.

Relationships

Read TEMPERA with rlprompt, autodspy, prompt-program-deployment-open-questions, and prompt-optimizer-regimes-for-harnesses. It is a useful contrast class for sammo and dspy, where optimization is more compile-time and artifact-oriented.