跳转至

朴素提示词优化：重新审视复杂提示词搜索的必要性

文章背景与核心概要

随着智能体 AI 的不断进步，寻找高效的自主智能体改进方法对于实现递归自我改进（RSI）至关重要。提示词优化作为一种替代模型权重微调的方案，因其在训练和部署阶段均具有较低的计算开销，且能提供相当的性能提升，正受到广泛关注。然而，近期研究往往倾向于开发极其复杂的提示词优化流程。

本文提出了“朴素提示词优化”（Naive Prompt Optimization, NPO），这是一种轻量级的单谱系方法，通过教师模型结合滚动反馈（rollout feedback）对提示词进行迭代优化。研究表明，NPO 在减少滚动次数的同时，能够达到甚至超越 GEPA 等复杂方法的性能。该研究的核心意义在于证明了简单的线性优化策略足以媲美复杂的搜索过程，且随着教师模型能力的增强，NPO 的优势愈发明显，为提示词工程提供了一种更高效、更具扩展性的范式。

📌 执行摘要

随着智能体 AI 的发展，高效改进自主智能体的方法对于递归自我改进（RSI）至关重要。提示词优化已成为权重微调的一种有前途的替代方案，它以更低的训练和部署计算开销提供了相当的性能提升。

然而，近期的研究越来越倾向于使用过于复杂的提示词优化程序。本文引入了朴素提示词优化（NPO），这是一种轻量级的单谱系方法，利用教师模型配合滚动反馈来迭代优化提示词。

本研究的主要结论包括： * 效率： NPO 在需要更少滚动次数的情况下，实现了与 GEPA 等复杂方法相当甚至更好的性能。 * 教师可扩展性： 当搭配更强大的教师模型时，NPO 的性能优势会进一步扩大，这表明更强的教师推理能力可以弥补优化器端搜索策略的复杂性。 * 任务通用性： 在交互式环境中，NPO 与 GEPA 相比仍具有竞争力，尽管某些不太适合提示词优化的任务可能更倾向于使用 GRPO 等方法。 * 可迁移性： 通过 NPO 优化的提示词在直接应用于其他学生模型时，能产生类似的性能提升，特别是在同一模型家族内部。

最终，研究结果表明，简单、线性的提示词优化可以有效地与极其复杂的搜索程序相抗衡。

As agentic AI advances, efficient methods to improve autonomous agents are critical for recursive self-improvement (RSI). Prompt optimization has emerged as a promising alternative to weight fine-tuning, offering comparable performance boosts with lower computational overhead for both training and deployment.

However, recent research increasingly gravitates toward overly complex prompt optimization procedures. This paper introduces Naive Prompt Optimization (NPO), a lightweight, single-lineage method that iteratively refines prompts using a teacher model paired with rollout feedback.

Key takeaways from the study include: * Efficiency: NPO achieves performance comparable to—or better than—more complex methods like GEPA, while requiring fewer rollouts. * Teacher Scalability: The performance advantage of NPO grows when paired with stronger teacher models, demonstrating that superior teacher reasoning can compensate for complex optimizer-side search strategies. * Task Versatility: In interactive environments, NPO remains competitive against GEPA, though certain tasks less suited for prompt optimization still favor methods like GRPO. * Transferability: Prompts optimized via NPO yield similar performance improvements when applied verbatim to other student models, particularly within the same model family.

Ultimately, the findings suggest that simple, linear prompt optimization can effectively rival significantly more intricate search procedures.

📄 摘要

高效地改进各类任务中的自主智能体，是加速智能体 AI 递归自我改进（RSI）的核心。提示词优化作为一种有前景的方法，能够提供与微调模型权重相当的性能增益，同时降低优化和推理阶段的计算成本。然而，近期的发展越来越倾向于使用不必要的复杂提示词优化器。我们引入了朴素提示词优化（NPO），这是一种轻量级的单谱系方法，通过教师模型结合滚动反馈来迭代修订提示词。NPO 在滚动次数更少的情况下，实现了与 GEPA 相当或更好的性能，且其优势随着教师模型能力的增强而增加，这表明更强的教师推理能力可以部分替代优化器端的搜索复杂性。在交互式游戏中，NPO 与 GEPA 相比保持了广泛的竞争力，而 GRPO 在某些不太适合提示词优化的任务上表现更好。我们还展示了 NPO 优化的提示词在直接应用于其他学生模型时，能产生类似的性能提升，特别是在同一模型家族内的模型之间。总体而言，我们的初步结果表明，简单、线性的提示词优化可以与实质上更复杂、更精密的搜索程序相媲美。

Efficiently improving autonomous agents across diverse tasks is central to accelerating recursive self-improvement (RSI) in agentic AI, with prompt optimization emerging as a promising approach capable of delivering performance gains comparable to those achieved by fine-tuning model weights, while reducing computational costs in both optimization and serving. However, recent developments increasingly favor unnecessarily complex prompt optimizers. We introduce Naive Prompt Optimization (NPO), a lightweight single-lineage method that iteratively revises prompts using a teacher model with rollout feedback. NPO achieves comparable or better performance than GEPA with fewer rollouts, and its advantage increases with stronger teacher models, suggesting that stronger teacher reasoning can partially substitute for optimizer-side search complexity. In interactive games, NPO remains broadly competitive with GEPA, while GRPO performs better on some tasks less amenable to prompt optimization. We also show that NPO-optimized prompts elicit similar performance improvements when applied verbatim to other student models, especially across models within the same family. Overall, our preliminary results show that simple, linear prompt optimization can rival substantially more sophisticated and complex search procedures.

🔗 快速链接与资源