---
title: "ZERA — paper by Yi et al. (2025) — summary and ideas"
description: "Automatic prompt optimization (APO) usually tunes only the user prompt. 27 references, cited by 4 papers here."
authors: ["Seungyoun Yi", "Minsoo Khang", "Sungrae Park"]
year: 2025
venue: "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing"
source: "https://aclanthology.org/2025.emnlp-main.1190/"
doi: "10.18653/v1/2025.emnlp-main.1190"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/yi-2025-zera-zero-init-instruction-evolving/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/yi-2025-zera-zero-init-instruction-evolving/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2025](https://md.research.elamur.ai/free/en/papers/2025/page/11/) / ZERA: Zero-init Instruction Evolving Refinement Agent – From Zero Instructions to Structured Prompts via Principle-based Optimization

# ZERA: Zero-init Instruction Evolving Refinement Agent – From Zero Instructions to Structured Prompts via Principle-based Optimization

**Authors:** Seungyoun Yi, Minsoo Khang and Sungrae Park · **Year:** 2025 · **Venue:** [Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing](https://md.research.elamur.ai/free/en/venues/emnlp/) ([EMNLP 2025](https://md.research.elamur.ai/free/en/venues/emnlp/2025/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2025.emnlp-main.1190/) · [PDF](https://aclanthology.org/2025.emnlp-main.1190.pdf) · [DOI 10.18653/v1/2025.emnlp-main.1190](https://doi.org/10.18653/v1/2025.emnlp-main.1190) · [OpenAlex](https://openalex.org/W4416035805)

**Citations:** 1 citations (OpenAlex, 2026-10-14) · [4](https://md.research.elamur.ai/free/en/papers/yi-2025-zera-zero-init-instruction-evolving/cited-by/) in our library, of the 149,439 papers whose bibliographies we read

## Summary
Automatic prompt optimization (APO) usually tunes only the user prompt. It learns from unstructured or metric-only feedback on a few examples and needs many samples and long iteration loops, which makes it costly and prone to overfitting. ZERA is a training-free agent that starts from essentially empty instructions and jointly evolves the system prompt, the user prompt and a task description. Each iteration has two stages:

- **Principle-based Critique Generation (PCG)** runs the task model on a small sample set. It infers how much each of eight general quality principles matters for the task, scores the outputs against them, and writes an analysis with suggestions.
- **Metacognitive Prompt Refinement (MPR)** rewrites the prompts using these critiques plus the history of earlier prompts.

ZERA was tested with five LLMs on nine benchmarks covering reasoning, summarization and code. It generally outperformed the default task prompts and the APO baselines PromptAgent, OPRO and CriSPO. Its value is in producing strong, interpretable prompts quickly from very little data, without hand-written templates. The code and all prompts are public (github.com/younatics/zera-agent).

## Key points
- **Principles:** The eight criteria span Reasoning Quality, Structural Alignment, Faithfulness, Correctness, Meaning Accuracy, Completeness, Expression Style and Conciseness. Their weights are inferred automatically per task, e.g. reasoning quality ranks highest for GSM8K and MMLU-Pro, while conciseness and faithfulness rank higher for CNN/DM and SAMSum.
- **Models and benchmarks:** GPT-3.5, GPT-4o, LLaMA-3.1-70B-Instruct, Qwen-2.5-70B-Instruct and Mistral-7B-Instruct-v0.3, evaluated on MMLU, MMLU-Pro, GSM8K, MBPP, HumanEval, BBH, HellaSwag, CNN/DM and SAMSum.
- **Principle ablation:** keeping only correctness and reasoning quality cuts BBH performance by 33.6 points compared with using all eight principles.
- **Prompt-component ablation:** using only the user prompt gives the lowest average score (≈0.559, vs ≈0.571 for full ZERA; values read from imperfectly extracted table text).
- **Transferability and limitations:**
  - Prompts tuned directly on LLaMA-3.1 beat prompts transferred from GPT-3.5 when run on LLaMA-3.1 (e.g. BBH 76.9 vs 72.9, MMLU-Pro 60.7 vs 57.3).
  - ZERA still needs about 5–20 samples per task.
  - Optimized prompts grow longer before stabilizing in length.
  - Summarization results rely only on ROUGE-L, without human evaluation.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

## Institutions

- [UpStage](https://md.research.elamur.ai/free/en/companies/upstage/) — Company, Yongin-si, South Korea · 6 works · 728 citations

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works

## Cite

Cite as: Seungyoun Yi, Minsoo Khang and Sungrae Park. 2025. ZERA: Zero-init Instruction Evolving Refinement Agent – From Zero Instructions to Structured Prompts via Principle-based Optimization. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{yi2025zera,
  title = {{ZERA: Zero-init Instruction Evolving Refinement Agent – From Zero Instructions to Structured Prompts via Principle-based Optimization}},
  author = {Seungyoun Yi, Minsoo Khang and Sungrae Park},
  booktitle = {Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
  doi = {10.18653/v1/2025.emnlp-main.1190},
  url = {https://doi.org/10.18653/v1/2025.emnlp-main.1190}
}
```

- [← previous Vulnerability of LLMs to Vertically Aligned Text Manipulations](https://md.research.elamur.ai/free/en/papers/li-2025-vulnerability-llms-vertically-aligned-text/)
- [next → A Survey of Machine Learning Models and Datasets for the Multi-label Classification of Textual Hate Speech in English](https://md.research.elamur.ai/free/en/papers/baumler-2025-survey-machine-learning-models-datasets/)
- [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/page/11/)

## References

27 references, 14 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, and 1 others. 2021. [Program synthesis with large language models](https://md.research.elamur.ai/free/en/papers/austin-2021-program-synthesis-large-language-models/). arXiv preprint arXiv:2108.07732.
2. Tom Brown, Benjamin Mann, Nick Ryder, and 1 others. 2020. [Language models are few-shot learners](https://md.research.elamur.ai/free/en/papers/brown-2020-language-models-few-shot-learners/). In NeurIPS.
3. Junhao Chen, Bowen Wang, Zhouqiang Jiang, and Yuta Nakashima. 2025. Putting people in llms’ shoes: Generating better answers via question rewriter. Proceedings of the AAAI Conference on Artificial Intelligence, 39(22):23577–23585. [find](https://scholar.google.com/scholar?q=Junhao+Chen%2C+Bowen+Wang%2C+Zhouqiang+Jiang%2C+and+Yuta+Nakashima%2E+2025%2E+Putting+people+in+llms%E2%80%99+shoes%3A+Generating+better+answers+via+question+rewriter%2E+Proceedings+of+the+AAAI+Conference+on+Artificial+Int)
4. Yuyan Chen, Zhihao Wen, Ge Fan, Zhengyu Chen, Wei Wu, Dayiheng Liu, Zhixu Li, Bang Liu, and Yanghua Xiao. 2024. Mapo: Boosting large language model performance with model-adaptive prompt optimization. arXiv preprint arXiv:2407.04118. [find](https://scholar.google.com/scholar?q=Yuyan+Chen%2C+Zhihao+Wen%2C+Ge+Fan%2C+Zhengyu+Chen%2C+Wei+Wu%2C+Dayiheng+Liu%2C+Zhixu+Li%2C+Bang+Liu%2C+and+Yanghua+Xiao%2E+2024%2E+Mapo%3A+Boosting+large+language+model+performance+with+model%2Dadaptive+prompt+optimization%2E)
5. Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Jared Kaplan, and 1 others. 2021. [Training verifiers to solve math word problems](https://md.research.elamur.ai/free/en/papers/cobbe-2021-training-verifiers-solve-math-word/). arXiv preprint arXiv:2110.14168.
6. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, and 1 others. 2024. [The llama 3 herd of models](https://md.research.elamur.ai/free/en/papers/grattafiori-2024-llama-3-herd-models/). arXiv preprint arXiv:2407.21783.
7. Bogdan Gliwa, Iwona Mochol, Michał Biesek, and Aleksander Wawer. 2019. [Samsum corpus: A human-annotated dialogue summary dataset](https://md.research.elamur.ai/free/en/papers/gliwa-2019-samsum-corpus-human-annotated-dialogue/). arXiv preprint arXiv:1911.12237.
8. Han He, Qianchu Liu, Lei Xu, Chaitanya Shivade, Yi Zhang, Sundararajan Srinivasan, and Katrin Kirchhoff. 2025. Crispo: Multi-aspect critique-suggestion-guided automatic prompt optimization for text generation. Proceedings of the AAAI Conference on Artificial Intelligence, 39(22):24014–24022. [find](https://scholar.google.com/scholar?q=Han+He%2C+Qianchu+Liu%2C+Lei+Xu%2C+Chaitanya+Shivade%2C+Yi+Zhang%2C+Sundararajan+Srinivasan%2C+and+Katrin+Kirchhoff%2E+2025%2E+Crispo%3A+Multi%2Daspect+critique%2Dsuggestion%2Dguided+automatic+prompt+optimization+for+text+ge)
9. Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Dawn Tang, Dawn Song, Jacob Steinhardt, and 1 others. 2021. [Measuring massive multitask language understanding](https://md.research.elamur.ai/free/en/papers/hendrycks-2020-measuring-massive-multitask-language-understanding/). arXiv preprint arXiv:2009.03300.
10. Karl Moritz Hermann, Tomás Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. [Teaching machines to read and comprehend](https://md.research.elamur.ai/free/en/papers/hermann-2015-teaching-machines-read-comprehend/). Advances in Neural Information Processing Systems, 28.
11. Yasaman Jafari, Dheeraj Mekala, Rose Yu, and Taylor Berg-Kirkpatrick. 2024. MORL-prompt: An empirical analysis of multi-objective reinforcement learning for discrete prompt optimization. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 9878–9889, Miami, Florida, USA. A… [find](https://scholar.google.com/scholar?q=Yasaman+Jafari%2C+Dheeraj+Mekala%2C+Rose+Yu%2C+and+Taylor+Berg%2DKirkpatrick%2E+2024%2E+MORL%2Dprompt%3A+An+empirical+analysis+of+multi%2Dobjective+reinforcement+learning+for+discrete+prompt+optimization%2E+In+Findings+o)
12. Zhen Jiang and 1 others. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825. [find](https://scholar.google.com/scholar?q=Zhen+Jiang+and+1+others%2E+2023%2E+Mistral+7b%2E+arXiv+preprint+arXiv%3A2310%2E06825%2E)
13. OpenAI. 2024. [Gpt-4o system card](https://md.research.elamur.ai/free/en/papers/openai-2024-gpt-4o-system-card/). arXiv preprint arXiv:2410.21276.
14. Dengyun Peng, Yuhang Zhou, Qiguang Chen, Jinhao Liu, Jingjing Chen, and Libo Qin. 2025. Dlpo: Towards a robust, efficient, and generalizable prompt optimization framework from a deep-learning perspective. arXiv preprint arXiv:2503.13413. [find](https://scholar.google.com/scholar?q=Dengyun+Peng%2C+Yuhang+Zhou%2C+Qiguang+Chen%2C+Jinhao+Liu%2C+Jingjing+Chen%2C+and+Libo+Qin%2E+2025%2E+Dlpo%3A+Towards+a+robust%2C+efficient%2C+and+generalizable+prompt+optimization+framework+from+a+deep%2Dlearning+perspect)
15. Ethan Perez and et al. 2021. [True few-shot learning with language models](https://md.research.elamur.ai/free/en/papers/perez-2021-true-few-shot-learning-language/). arXiv preprint arXiv:2105.11447.
16. Reid Pryzant, Dan Iter, Jerry Li, Yin Lee, Chenguang Zhu, and Michael Zeng. 2023. [Automatic prompt optimization with “gradient descent” and beam search](https://md.research.elamur.ai/free/en/papers/pryzant-2023-automatic-prompt-optimization-gradient-descent/). In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7957–7968, Singapore. Association for Computationa…
17. Saurabh Srivastava and Ziyu Yao. 2025. Revisiting prompt optimization with large reasoning models-a case study on event extraction. arXiv preprint arXiv:2504.07357. — [arXiv:2504.07357](https://arxiv.org/abs/2504.07357) [find](https://scholar.google.com/scholar?q=Saurabh+Srivastava+and+Ziyu+Yao%2E+2025%2E+Revisiting+prompt+optimization+with+large+reasoning+models%2Da+case+study+on+event+extraction%2E+arXiv+preprint+arXiv%3A2504%2E07357%2E+%E2%80%94+%5BarXiv%3A2504%2E07357%5D%28https%3A%2F%2Farxiv%2E)
18. Mirac Suzgun, Nathan Scales, Nathanael Schärli, Aitor Lewkowycz, Mikael Stenmark, Shunyu Yao, Adams Yu, Jacob Austin, Aakanksha Chowdhery, Quoc Le, and 1 others. 2022. [Challenging big-bench tasks and whether chain-of-thought can solve them](https://md.research.elamur.ai/free/en/papers/suzgun-2023-challenging-big-bench-tasks-whether/). arXiv preprint arXiv:2210.09261.
19. Qwen Team. 2024. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115. [find](https://scholar.google.com/scholar?q=Qwen+Team%2E+2024%2E+Qwen2%2E5+technical+report%2E+arXiv+preprint+arXiv%3A2412%2E15115%2E)
20. Xinyuan Wang, Chenxi Li, Zhen Wang, Fan Bai, Haotian Luo, Jiayou Zhang, Nebojsa Jojic, Eric Xing, and Zhiting Hu. 2024. Promptagent: Strategic planning with language models enables expert-level prompt optimization. In The Twelfth International Conference on Learning Representations. [find](https://scholar.google.com/scholar?q=Xinyuan+Wang%2C+Chenxi+Li%2C+Zhen+Wang%2C+Fan+Bai%2C+Haotian+Luo%2C+Jiayou+Zhang%2C+Nebojsa+Jojic%2C+Eric+Xing%2C+and+Zhiting+Hu%2E+2024%2E+Promptagent%3A+Strategic+planning+with+language+models+enables+expert%2Dlevel+prompt)
21. Jinyu Xiang, Jiayi Zhang, Zhaoyang Yu, Fengwei Teng, Jinhao Tu, Xinbing Liang, Sirui Hong, Chenglin Wu, and Yuyu Luo. 2025. Self-supervised prompt optimization. arXiv preprint arXiv:2502.06855. [find](https://scholar.google.com/scholar?q=Jinyu+Xiang%2C+Jiayi+Zhang%2C+Zhaoyang+Yu%2C+Fengwei+Teng%2C+Jinhao+Tu%2C+Xinbing+Liang%2C+Sirui+Hong%2C+Chenglin+Wu%2C+and+Yuyu+Luo%2E+2025%2E+Self%2Dsupervised+prompt+optimization%2E+arXiv+preprint+arXiv%3A2502%2E06855%2E)
22. Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V Le, Denny Zhou, and Xinyun Chen. 2024. [Large language models as optimizers](https://md.research.elamur.ai/free/en/papers/yang-2023-large-language-models-optimizers/). In The Twelfth International Conference on Learning Representations.
23. Junjie Ye, Xuanting Chen, Nuo Xu, Can Zu, Zekai Shao, Shichun Liu, Yuhan Cui, Zeyang Zhou, Chao Gong, Yang Shen, and 1 others. 2023. A comprehensive capability analysis of gpt-3 and gpt-3.5 series models. arXiv preprint arXiv:2303.10420. [find](https://scholar.google.com/scholar?q=Junjie+Ye%2C+Xuanting+Chen%2C+Nuo+Xu%2C+Can+Zu%2C+Zekai+Shao%2C+Shichun+Liu%2C+Yuhan+Cui%2C+Zeyang+Zhou%2C+Chao+Gong%2C+Yang+Shen%2C+and+1+others%2E+2023%2E+A+comprehensive+capability+analysis+of+gpt%2D3+and+gpt%2D3%2E5+series+mod)
24. Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. Hellaswag: Can a machine really finish your sentence? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4791–4800. [find](https://scholar.google.com/scholar?q=Rowan+Zellers%2C+Ari+Holtzman%2C+Yonatan+Bisk%2C+Ali+Farhadi%2C+and+Yejin+Choi%2E+2019%2E+Hellaswag%3A+Can+a+machine+really+finish+your+sentence%3F+In+Proceedings+of+the+57th+Annual+Meeting+of+the+Association+for+Com)
25. Yuxiang Zhang and Jitao Sang. 2025. Encoder of thoughts: Enhancing planning ability in language agents through structural embedding. Proceedings of the AAAI Conference on Artificial Intelligence, 39(24):25994–26002. [find](https://scholar.google.com/scholar?q=Yuxiang+Zhang+and+Jitao+Sang%2E+2025%2E+Encoder+of+thoughts%3A+Enhancing+planning+ability+in+language+agents+through+structural+embedding%2E+Proceedings+of+the+AAAI+Conference+on+Artificial+Intelligence%2C+39%282)
26. Zhengxuan Zhao, Eric Wallace, Shi Wang, and et al. 2021. [Calibrate before use: Improving few-shot performance of language models](https://md.research.elamur.ai/free/en/papers/zhao-2021-calibrate-before-use-improving-few/). In ICML.
27. Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. 2023. [Large language models are human-level prompt engineers](https://md.research.elamur.ai/free/en/papers/zhou-2022-large-language-models-human-level/). In The Eleventh International Conference on Learning Representations.

*Updated 30 October 2026*