Title: SePO: Self-Evolving Prompt Agent for System Prompt Optimization

Abstract

Optimizing system prompts enhances agent performance without altering the underlying model architecture, resulting in instructions that are both model-agnostic and interpretable by humans. While current approaches utilize a prompt agent to refine the system prompts of task-specific agents, the system prompt governing the prompt agent itself is typically hand-crafted and static. To address this limitation, we introduce Self-Evolving Prompt Optimization (SePO), a framework that designates the prompt agent’s own system prompt as an optimization objective alongside those of the task agents.

SePO employs a self-referential architecture. Leveraging an open-ended evolutionary search process that preserves an archive of candidate prompts as evolutionary milestones, a single prompt agent simultaneously optimizes its own system prompt and those of the task agents. The training methodology consists of two phases: pre-training, where the prompt agent evolves across a diverse multi-task pool, and fine-tuning, which adapts the agent to a specific target task.

We evaluated SePO across five distinct benchmarks: mathematics (AIME'25), abstract reasoning (ARC-AGI-1), graduate-level science (GPQA), code generation (MBPP), and logic puzzles (Sudoku). In these tests, SePO consistently surpassed existing methods such as Manual-CoT, TextGrad, and MetaSPO, achieving an average accuracy gain of 4.49 points over Manual-CoT. Furthermore, the prompt optimization capabilities acquired during pre-training demonstrated strong generalization to tasks outside the pre-training distribution, indicating that the model learns transferable skills rather than merely memorizing task-specific prompts.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC