RLMOpt: Adaptive Prompt Optimization via Recursive Language Models

Prompt optimization automates the process of refining language model prompts to improve task performance. Traditional optimizers rely on fixed, predefined search procedures—such as evolutionary algorithms or Bayesian optimization—to generate and evaluate candidate prompts. RLMOpt introduces a paradigm shift by using a recursive language model (RLM) as an adaptive agent that controls the search process itself. This agent dynamically decides when to inspect task data, analyze failure patterns, generate new candidates, and allocate the evaluation budget.

RLMOpt separates the search policy from the evaluation process. The RLM agent acts as the decision-maker, while a deterministic harness handles the objective evaluation. This division of labor is critical: the agent has the flexibility to explore complex search strategies, but the harness ensures that all candidates are rigorously evaluated against objective metrics. The harness enforces Pareto-based selection and regression constraints, preventing the optimizer from sacrificing performance in one area to gain it in another. This structure allows the system to remain reliable even when evaluation data is noisy or limited.

In head-to-head comparisons across four benchmarks—including clinical information extraction, multi-hop question answering, instruction following, and tool-calling agents—RLMOpt consistently outperformed the baseline optimizer (GEPA). It achieved higher held-out scores in 9 out of 11 matched benchmark-seed comparisons. Furthermore, RLMOpt proved more efficient, requiring fewer search rollouts to reach its peak performance and producing significantly more concise prompts. The study highlights that the ultimate success of prompt optimization is bounded by the 'headroom' of the initial seed prompt; the optimizer's role is to reliably reach that ceiling with minimal computational cost.