SePO: Self-Evolving Prompt Agent for System Prompt Optimization
TL;DR. System-prompt optimization improves agent behavior without touching model weights, yielding human-readable, model-agnostic instructions. Existing methods build a prompt agent that refines task agents' prompts but leave the prompt agent's own system prompt hand-engineered and fixed. SePO makes the prompt agent's own prompt an optimization target too, in a self-referential design: one prompt agent improves both task agents' prompts and its own, under an open-ended evolutionary search that keeps an archive of candidate prompts as stepping stones. It pre-trains on a multi-task pool, then fine-tunes on a target task. Across five benchmarks (AIME'25, ARC-AGI-1, GPQA, MBPP, Sudoku) it beats Manual-CoT, TextGrad, and MetaSPO, +4.49 average points over Manual-CoT, and the learned skill generalizes beyond the pre-training mixture rather than memorizing per-task prompts.
Source: HuggingFace Daily Papers (upvotes: 1) arxiv: 2606.04465 Raw: raw/huggingface/2026-06-05-sepo-self-evolving-prompt-agent-for-system-prompt-optimizati.md
Key points
- Self-reference: the prompt agent optimizes its own system prompt alongside the task agents' prompts, closing the loop that prior methods left hand-engineered.
- Open-ended evolutionary search with a candidate archive used as stepping stones (a quality-diversity flavor), rather than greedy refinement.
- Two-stage: pre-train the prompt-optimization skill on a multi-task pool, then fine-tune on the target task.
- Generalization: the pre-trained skill transfers to tasks outside the pre-training mixture, so it learns how to optimize prompts rather than memorizing specific prompts, the same principle-vs-instance distinction the cluster keystone draws.
Relation to prior wiki
Part of today's self-evolving-agents cluster (Continual Experience Internalization). SePO is the prompt-space instance of "make the optimizer self-improving," and its finding that the skill generalizes while per-task prompts would not is the exact principle-level-beats-instance-level result the keystone paper formalizes. It also rhymes with the broader self-referential trend the wiki has tracked in self-evolving environments (EvoEnv, 05-15): the system evolves the thing that shapes its own learning.