RLMOpt Uses Recursive LLMs for Adaptive Prompt Optimization

This paper introduces RLMOpt, an innovative prompt optimizer that significantly advances the field of automated prompt engineering. Unlike conventional methods that rely on predefined optimization procedures, RLMOpt empowers a recursive language model (RLM) agent to control the search policy itself. This RLM agent operates within a tool-based environment, enabling it to inspect task information, analyze failures, generate new prompt candidates, intelligently allocate evaluation budgets, and decide when to conclude the optimization process. A deterministic harness complements the RLM agent, ensuring objective scoring, Pareto-based selection, and adherence to regression constraints. The effectiveness of RLMOpt was rigorously evaluated across four diverse benchmarks, including clinical information extraction, multi-hop question answering, verifiable instruction following, and multi-turn tool-calling agents. In comparative tests, RLMOpt consistently achieved the best held-out scores across all benchmarks and outperformed a leading baseline (GEPA) in most matched comparisons. It also demonstrated superior efficiency, achieving these results with fewer search rollouts and producing more concise prompts. The research highlights that optimization gains are primarily determined by the initial prompt's headroom, emphasizing the importance of reliably reaching this potential with minimal search effort.