SePO: Self-Evolving Prompt Agent for System Prompt Optimization
2026-06-03 • Computation and Language
Computation and LanguageArtificial Intelligence
AI summaryⓘ
The authors propose a new method called Self-Evolving Prompt Optimization (SePO) that not only improves instructions given to task agents but also improves its own instructions automatically. Unlike previous methods where the prompt agent's own instructions were fixed, SePO evolves both its own and task agents' prompts through a continuous search process. They train this approach in two steps: first on many tasks, then fine-tune on a specific task. Their experiments on five different types of problems show SePO performs better than other prompt optimization methods, and its training helps it generalize to new tasks instead of just memorizing prompts.
system prompt optimizationprompt agentself-referential designevolutionary searchmulti-task trainingfine-tuningtask agentsmodel-agnostic instructionsprompt generalizationbenchmark evaluation
Authors
Wangcheng Tao, Han Wu, Weng-Fai Wong
Abstract
System prompt optimization improves agent behavior without modifying the underlying model, yielding human-readable, model-agnostic instructions. Existing methods build a prompt agent that refines task agents' system prompts, yet leave the prompt agent's own system prompt hand-engineered and fixed. We propose Self-Evolving Prompt Optimization (SePO), which treats the prompt agent's own system prompt as an optimization target alongside task agents' system prompts. SePO adopts a self-referential design. A single prompt agent improves both task agents' system prompts and its own under an open-ended evolutionary search that maintains an archive of candidate prompts as stepping stones. Training proceeds in two stages: pre-training evolves the prompt agent on a multi-task pool, and fine-tuning then applies it to a target task. Across five benchmarks spanning math (AIME'25), abstract reasoning (ARC-AGI-1), graduate-level science (GPQA), code generation (MBPP), and logic puzzles (Sudoku), SePO consistently outperforms Manual-CoT, TextGrad, and MetaSPO, improving the average accuracy by 4.49 points compared to Manual-CoT. The prompt optimization skill from pre-training also generalizes to tasks beyond the pre-training mixture, rather than memorizing per-task prompts.