SePO: Self-Evolving Prompt Agent for System Prompt Optimization

Researchers propose Self-Evolving Prompt Optimization (SePO), a novel system that automatically optimizes AI agent prompts by treating the prompt agent's own instructions as an optimization target. The method demonstrates consistent performance gains across five diverse benchmarks, outperforming existing approaches and showing generalization to unseen tasks.

SePO represents a meaningful advancement in prompt engineering by introducing self-referential optimization—a prompt agent that improves both task-specific prompts and refines its own instructions iteratively. This addresses a critical limitation in existing prompt optimization methods that rely on hand-engineered, static instructions for the optimizer itself. The self-evolving architecture maintains an archive of candidate prompts as exploration stepping stones, enabling open-ended discovery of increasingly effective instructions.

The research builds on growing recognition that system prompts significantly influence AI agent behavior without requiring model retraining. Prior work like TextGrad and MetaSPO demonstrated that automated prompt refinement could improve performance, yet left the optimization process itself unoptimized. SePO closes this gap through a two-stage training pipeline combining pre-training across diverse tasks and fine-tuning on target problems. Results spanning mathematics (AIME'25), abstract reasoning (ARC-AGI-1), science (GPQA), code generation (MBPP), and logic puzzles show consistent 4.49-point accuracy improvements over baseline CoT approaches.

The generalization findings carry particular importance—the prompt optimization skill transfers to tasks beyond pre-training distribution, indicating the method learns genuine optimization strategies rather than memorizing task-specific solutions. This portability makes SePO potentially valuable for practitioners seeking model-agnostic improvements across heterogeneous workloads. The human-readable output maintains interpretability advantages over gradient-based optimization. For AI infrastructure developers and enterprises deploying multi-agent systems, this work suggests substantial performance gains are achievable through instruction refinement alone, reducing computational overhead compared to model fine-tuning approaches.

• →Self-referential prompt optimization enables prompt agents to improve their own instructions while optimizing task-specific prompts, closing a gap in existing methods.
• →SePO demonstrates 4.49-point average accuracy improvements across five diverse benchmarks spanning math, reasoning, science, coding, and logic domains.
• →The two-stage training approach (pre-training on task pools followed by fine-tuning) enables optimization skills to generalize to unseen tasks beyond pre-training distribution.
• →Model-agnostic prompt optimization preserves interpretability and avoids retraining costs, making it practical for deploying improved agent behavior across existing systems.
• →Archive-based evolutionary search with stepping stones provides a scalable mechanism for open-ended prompt discovery without gradient computations.