Prompting Policies: RL-Trained Prompter Lifts Black-Box LLM Reasoning from 55% to 90%

arXiv·high signal

Google Research introduces a reinforcement learning framework that trains a lightweight prompter model to maximize task-specific rewards for a larger frozen worker LLM, using a contrastive experience buffer coupling scalar rewards with dense textual critiques. On Big Bench Extra Hard (BBEH) and Tau-bench, the approach improves logic-intensive reasoning from 55% to 90% and tool-use tasks from 74% to 91%. Key insight: prompt engineering can be amortized into learned single-shot policy weights rather than requiring iterative manual refinement.