prompt

Free

Self-evolving prompt agent for open-ended system prompt optimization

FreeFree tier

Type

Open Source

Visit Website GitHub (40 stars)

About prompt

SePO (Self-Evolving Prompt Agent) is a self-referential system prompt optimization method that treats the optimizer's own system prompt as an optimization target rather than a fixed artifact. It employs an open-ended evolutionary search with an archive of candidate prompts used as stepping stones, and a two-stage pipeline: pre-training on a diverse multi-task pool to learn a transferable prompt-optimization skill, followed by fine-tuning on a specific target task. Based on research from NUS and CityUHK (arXiv 2606.04465, June 2026), SePO demonstrates generalization to held-out tasks such as Sudoku and achieves an average of +4.49 points over Manual-CoT on benchmarks including AIME'25, ARC-AGI-1, GPQA, and MBPP.

Key Features

Self-referential evolution: the prompt agent's own system prompt is also an optimization target, not a fixed constant.

Two-stage pipeline: pre-train on a multi-task pool to learn transferable skills, then fine-tune on a target task.

Archive as stepping stones: maintains an archive of candidate prompts used for future mutations and selection.

Open-ended evolutionary search: generates new candidates through mutation (rephrase, add constraints, examples, etc.) and selects based on measured performance.

Generalization to held-out tasks: works on tasks not seen during pre-training, such as Sudoku.

Improved benchmark performance: reported average +4.49 points over Manual-CoT on AIME'25, ARC-AGI-1, GPQA, MBPP, and Sudoku.

Pros & Cons

Pros

Self-improving: the optimizer can evolve its own system prompt, closing the optimization loop.
Generalizes well to unseen tasks, reducing need for per-task manual prompt engineering.
Open-source and freely available for use and modification.
Backed by published research and benchmark results showing significant gains over Manual-CoT.
Archive mechanism preserves partial solutions, enabling exploration of diverse prompts.

Cons

Not a standalone tool; requires integration with an LLM and implementation of the evolutionary loop.
Computational cost may be high due to repeated evaluation of candidate prompts during evolution.
Documentation is limited to the research paper and the prompt file; no official package or demos provided.

Best For

Optimizing system prompts for large language models across diverse reasoning tasks.Automating prompt engineering for benchmarks like AIME, ARC-AGI, GPQA, MBPP, and Sudoku.Research in self-improving AI agents and meta-learning for prompt optimization.Developing transferable prompt optimization skills that can be specialized per domain.

FAQ

What is SePO?

SePO (Self-Evolving Prompt Agent) is a method for optimizing system prompts through a self-referential evolutionary loop. It treats the optimizer's own system prompt as a parameter to be evolved, uses an archive of candidate prompts as stepping stones, and operates in a two-stage pipeline: pre-training on a multi-task pool to learn a transferable skill, then fine-tuning on a target task. It was introduced in a June 2026 arXiv paper from NUS and CityUHK.

How does SePO differ from standard prompt optimization?

Standard prompt optimization treats the optimizer's system prompt as a fixed hand-engineered constant. SePO closes the loop by evolving the optimizer's own prompt alongside task agents' prompts, allowing self-improvement. It also uses an archive of candidate prompts (stepping stones) rather than a single best prompt, and pre-trains on multiple tasks before fine-tuning on a specific one.

What benchmarks has SePO been tested on?

SePO has been evaluated on AIME'25, ARC-AGI-1, GPQA, MBPP, and Sudoku, achieving an average of +4.49 points over the Manual-CoT baseline.