System Prompt Optimization with Meta-Learning (2505.09666v1)

Abstract: LLMs have shown remarkable capabilities, with optimizing their input prompts playing a pivotal role in maximizing their performance. However, while LLM prompts consist of both the task-agnostic system prompts and task-specific user prompts, existing work on prompt optimization has focused on user prompts specific to individual queries or tasks, and largely overlooked the system prompt that is, once optimized, applicable across different tasks and domains. Motivated by this, we introduce the novel problem of bilevel system prompt optimization, whose objective is to design system prompts that are robust to diverse user prompts and transferable to unseen tasks. To tackle this problem, we then propose a meta-learning framework, which meta-learns the system prompt by optimizing it over various user prompts across multiple datasets, while simultaneously updating the user prompts in an iterative manner to ensure synergy between them. We conduct experiments on 14 unseen datasets spanning 5 different domains, on which we show that our approach produces system prompts that generalize effectively to diverse user prompts. Also, our findings reveal that the optimized system prompt enables rapid adaptation even to unseen tasks, requiring fewer optimization steps for test-time user prompts while achieving improved performance.

System Prompt Optimization with Meta-Learning

In the paper titled "System Prompt Optimization with Meta-Learning," the authors tackle an underexplored aspect of optimizing LLMs by focusing on system prompts. Unlike user prompts which are tailored for specific tasks, system prompts are task-agnostic and have the potential to guide LLM behavior across diverse tasks. The paper introduces a novel formulation of bilevel system prompt optimization, where the system prompt serves as a higher-level optimization target to enhance the model's performance on various user prompts and tasks. This hierarchical problem is addressed through a meta-learning framework, termed MetaSPO (Meta-level System Prompt Optimizer), aiming to generalize prompts beyond the optimization environment.

The primary contribution of this work is the introduction of a meta-learning framework to solve the optimization problem presented by system prompts. MetaSPO involves alternating optimization loops: the inner loop focuses on refining user prompts, while the outer loop optimizes the system prompt by aligning it with the varied set of user prompts optimized in the inner loop. This iterative process is designed to ensure the robustness and generalizability of system prompts across multiple domains and tasks.

Empirical validation is extensive, spanning 14 unseen datasets across five distinct domains. The authors explore two real-world scenarios: unseen generalization, where system prompts are applied without further optimization, and test-time adaptation, where user prompts are further refined with few available examples from target tasks. In both scenarios, MetaSPO notably outperforms baseline methods including SPRIG and other hand-crafted strategies. Notably, MetaSPO achieves an average performance improvement of over 9% compared to default methods in unseen generalization setups.

MetaSPO's design is advantageous in settings lacking abundant task-specific data for prompt optimization. The authors demonstrate that optimized system prompts can generalize effectively even to dissimilar tasks and domains, highlighting the framework's adaptability and efficiency. Additionally, they show that in scenarios requiring rapid adaptation, MetaSPO facilitates convergence and improves performance in fewer steps than those needed by traditional methods, proving both efficient and cost-effective.

The paper notes positive correlations between source-target task similarity and performance improvements when using optimized prompts, suggesting that selecting source tasks closer to the target domain enhances the framework's efficacy. However, even tasks with lower similarity still benefit significantly from MetaSPO's system prompt optimization, indicating robust transferability.

Future work can explore integrating MetaSPO with even more diverse datasets and exploring its application with smaller, less capable optimization models to test latent application potentials. Additionally, protecting against potential misuse in unethical contexts remains a vital consideration in deploying MetaSPO at scale.

In conclusion, this research advances the understanding and practical application of system prompt optimization in LLMs. By leveraging meta-learning, it provides nuanced insights into achieving generalizable, efficient, and adaptable LLMs useful across a wide array of domains and contexts. As AI continually evolves, such frameworks will be essential for optimizing and guiding the behavior of increasingly sophisticated models.