TEMPERA: Test-Time Prompting via Reinforcement Learning

Source: arXiv Authors: Tianjun Zhang, Xuezhi Wang, Denny Zhou, Dale Schuurmans, Joseph E. Gonzalez Date: 2022-11-21 Primary category: cs.CL All categories: cs.CL, cs.AI

Abstract / key passage

TEMPERA trains a reinforcement-learning policy to edit an initial prompt at inference time for each query. Its action space spans commonly used prompt components such as instruction phrases, few-shot exemplars, and verbalizers, so the method is runtime prompt adaptation rather than one-shot offline prompt search.

Harness takeaway

TEMPERA is a useful anchor for the runtime-adaptation branch of the literature. It shows that prompt surfaces can be treated as live control surfaces per query without updating model weights, though the learned policy is still trained offline and the setting is much narrower than long-horizon agent workflows.