Elamur Research⌕
☰
home / papers / 2026 / Adaptive Prompt Optimization for Open-Ended Tasks: Uncertainty Preference as a Secondary Signal

Adaptive Prompt Optimization for Open-Ended Tasks: Uncertainty Preference as a Secondary Signal

Paper · 2026 · Shuyang Zhang, Zhixuan Liu, Zhichen Dong et al.

Summary

Training-free prompt optimizers usually rank candidate prompts by a single quality score. This ignores that tasks differ in how much output variability they want. Safety and factual tasks benefit from stable, predictable answers, while creative tasks benefit from diversity. The authors first test this idea with hand-written conservative and creative guidance on safety, factual-coherence and open-ended writing benchmarks. They track both task scores and a semantic-entropy proxy, computed by clustering several sampled responses by meaning.

They then propose style-probing prompt evolution. Lightweight probes infer whether a task favours low or high entropy, and each candidate's judge score is adjusted up or down depending on whether its entropy matches that preference. The method plugs into OPRO and SPO and needs no training or access to model internals.

On MT-Bench subsets, the add-on modestly but fairly consistently raises pairwise win rates. Entropy also moves in the direction each subset prefers. This makes uncertainty a cheap secondary signal for black-box prompt tuning.

Key points

  • Diagnostic study: 21 of 24 model × task combinations showed the expected pattern.
    • Benchmarks: SALAD-Bench, In-the-wild Jailbreak, WikiText, WikiNews, NoveltyBench and EQ-Bench.
    • Models: Llama3.2-3B/3.1-8B-Instruct and Qwen2.5-1.5B/7B-Instruct.
    • Conservative prompts helped safety and coherence tasks; creative prompts helped novelty and EQ tasks.
  • Score shaping: the raw judge score (0–100) gets an extra term, α × preference sign × (normalized semantic entropy − threshold). The preference sign is +1, 0 or −1 and is inferred via Pearson correlation between probe scores and style labels.
  • Optimization results: the method was tested on MT-Bench Writing, Roleplay, Humanities and STEM, with GPT-4o as judge and 4 optimization rounds, using Llama3.1, Qwen3-4B and Qwen2.5-7B.
    • Adding the method to OPRO/SPO improved 48 of 72 cells, tied in 9 and dropped in 15.
    • The average gain was about +0.42 points.
  • Entropy shifts matched inferred preferences: semantic entropy rose for Writing (0.44 → 0.52) and Roleplay, and fell for STEM (0.27 → 0.20).
  • Limitations:
    • Repeated sampling raises inference cost.
    • LLM-judge biases may affect both optimization and evaluation.
    • The method could be misused for unsafe objectives.
    • Code is available at github.com/syzonline/Prompt-Evolving.

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

🎓Shanghai Jiao Tong UniversityShanghai · China212 works · 19,544 citations🎓University of Science and Technology of ChinaHefei · China145 works · 114,409 citations🔬Shanghai Artificial Intelligence LaboratoryShanghai · China152 works · 13,166 citations🎓University of Chinese Academy of SciencesBeijing · China188 works · 8,743 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 works

🧭 Topics 1Research subjects and languages the work is about.Bias, fairness & safety · 587 works

Cite

Cite as Shuyang Zhang, Zhixuan Liu, Zhichen Dong, Hao Zhang, Chaochao Lu and Chao Yang. 2026. Adaptive Prompt Optimization for Open-Ended Tasks: Uncertainty Preference as a Secondary Signal. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.1692

@inproceedings{zhang2026adaptive,
  title = {{Adaptive Prompt Optimization for Open-Ended Tasks: Uncertainty Preference as a Secondary Signal}},
  author = {Shuyang Zhang, Zhixuan Liu, Zhichen Dong, Hao Zhang, Chaochao Lu and Chao Yang},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1692},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1692}
}

References

33 references, 17 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. self-improvement via prompt evolution. In NeurIPS. ArXiv:2309.16797. — Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution
  2. Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen, Shengjie Ma, Honghao Liu, Saizhuo Wang, Kun Zhang, Yuanzhuo Wang, Wen Gao, Lionel Ni, and Jian Guo. 2025. A survey on llm-as-a-judge. Preprint, arXiv:2411.15594. find
  3. Qingfu Guo and 1 others. 2023. Evoprompt: Connecting large language models with evolutionary algorithms yields powerful prompt optimizers. arXiv preprint arXiv:2309.08532.
  4. Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. 2023. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. Preprint, arXiv:2302.09664.
  5. Lijun Li, Bowen Dong, Ruohui Wang, Xuhao Hu, Wangmeng Zuo, Dahua Lin, Yu Qiao, and Jing Shao. 2024. Salad-bench: A hierarchical and comprehensive safety benchmark for large language models. arXiv preprint arXiv:2402.05044.
  6. Zhen Lin, Shubhendu Trivedi, and Jimeng Sun. 2024. Generating with confidence: Uncertainty quantification for black-box large language models. Preprint, arXiv:2305.19187. find
  7. Chris Yuhao Liu, Liang Zeng, Jiacai Liu, Rui Yan, Jujie He, Chaojie Wang, Shuicheng Yan, Yang Liu, and Yahui Zhou. 2024. Skywork-reward: Bag of tricks for reward modeling in llms. arXiv preprint arXiv:2410.18451.
  8. Abhinav Jauhri et.al Aaron Grattafiori, Abhimanyu Dubey. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.
  9. Wenxin Luo, Weirui Wang, Xiaopeng Li, Weibo Zhou, Pengyue Jia, and Xiangyu Zhao. 2025. Tapo: Task-referenced adaptation for prompt optimization. Preprint, arXiv:2501.06689. — [arXiv:2501.06689](https://arxiv.org/abs/2501.06689) find
  10. Ekansh Agarwal and 1 others. 2024. Promptwizard: Task-aware prompt optimization framework. arXiv preprint arXiv:2405.18369. — [arXiv:2405.18369](https://arxiv.org/abs/2405.18369) find
  11. Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. 2016. Pointer sentinel mixture models. Preprint, arXiv:1609.07843.
  12. Kushal Arora, Timothy J. O’Donnell, Doina Precup, Jason Weston, and Jackie C. K. Cheung. 2023. The stable entropy hypothesis and entropy-aware decoding: An analysis and algorithm for robust natural language generation. Preprint, arXiv:2302.06784. — [arXiv:2302.06784](https://arxiv.org/abs/2302.06784) find
  13. OpenAI. 2024. Gpt-4 technical report. arXiv:2303.08774.
  14. Wei-Lin Chiang, Lianmin Zheng, Ying Sheng, Anastasios Nikolas Angelopoulos, Tianle Li, Dacheng Li, Hao Zhang, Banghua Zhu, Michael Jordan, Joseph E. Gonzalez, and Ion Stoica. 2024. Chatbot arena: An open platform for evaluating llms by human preference. Preprint, arXiv:2403.04132.
  15. Ganqu Cui, Yuchen Zhang, Jiacheng Chen, Lifan Yuan, Zhi Wang, Yuxin Zuo, Haozhan Li, Yuchen Fan, Huayu Chen, Weize Chen, Zhiyuan Liu, Hao Peng, Lei Bai, Wanli Ouyang, Yu Cheng, Bowen Zhou, and Ning Ding. 2025. The entropy mechanism of reinforcement learning for reasoning language models. Preprint, a…
  16. Chrisantha Fernando, Dylan Banarse, Henryk Michalewski, Simon Osindero, and Tim Rocktäschel. 2023. find
  17. Promptbreeder: Self-referential Preprint, Samuel J Paech. 2023. Eq-bench: An emotional intelligence benchmark for large language models. arXiv preprint arXiv:2312.06281.
  18. Samuel J Paech. 2025. Eq-bench creative writing benchmark v3. https://github.com/EQ-bench/ creative-writing-bench. ↗
  19. Max Peeperkorn, Tom Kouwenhoven, Dan Brown, and Anna Jordanous. 2024. Is temperature the creativity parameter of large language models? Preprint, arXiv:2405.00492. — [arXiv:2405.00492](https://arxiv.org/abs/2405.00492) find
  20. Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, and 25 others. 2025. Qwen2.5 technical report. Preprint, arXiv:2412.15115. find
  21. Xinyue Shen, Zeyuan Chen, Michael Backes, Yun Shen, and Yang Zhang. 2024. "do anything now": Characterizing and evaluating in-the-wild jailbreak prompts on large language models. In Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security, CCS ’24, page 1671–1685, New…
  22. Sridevi Wagle, Sai Munikoti, Anurag Acharya, Sara Smith, and Sameera Horawalavithana. 2023. Empirical evaluation of uncertainty quantification in retrieval-augmented language models for science. Preprint, arXiv:2311.09358. 2023. Large language models are human-level prompt engineers. Preprint, arXiv…
  23. Daniel M. Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B. Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. 2020. Fine-tuning language models from human preferences. Preprint, arXiv:1909.08593. find
  24. Yikun Wang, Rui Zheng, Liang Ding, Qi Zhang, Dahua Lin, and Dacheng Tao. 2024. Uncertainty aware learning for language model alignment. Preprint, arXiv:2406.04854. find
  25. Chuan Wu, Evangelos Kanoulas, Maarten de Rijke, and Wei Lu. 2020. WN-salience: A corpus of news articles with entity salience annotations. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 2095–2102, Marseille, France. European Language Resources Association.
  26. Jialiang Wu, Yi Shen, Sijia Liu, Yi Tang, Sen Song, Xiaoyi Wang, and Longjun Cai. 2025. Improve decoding factuality by token-wise cross layer entropy of large language models. arXiv preprint arXiv:2502.03199. find
  27. Jinyu Xiang, Jiayi Zhang, Zhaoyang Yu, Xinbing Liang, Fengwei Teng, Jinhao Tu, Fashen Ren, Xiangru Tang, Sirui Hong, Chenglin Wu, and Yuyu Luo. 2025. Self-supervised prompt optimization. Preprint, arXiv:2502.06855. find
  28. Sang Michael Xie, Aditi Raghunathan, Percy Liang, and Tengyu Ma. 2021. An explanation of in-context learning as implicit bayesian inference. arXiv preprint arXiv:2111.02080.
  29. Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V. Le, Denny Zhou, and Xinyun Chen. 2023. Large language models as optimizers. arXiv preprint arXiv:2309.03409.
  30. Yiming Zhang, Harshita Diddee, Susan Holm, Hanchen Liu, Xinyue Liu, Vinay Samuel, Barry Wang, and Daphne Ippolito. 2025a. Noveltybench: Evaluating language models for humanlike diversity. arXiv preprint arXiv:2504.05228. — [arXiv:2504.05228](https://arxiv.org/abs/2504.05228) find
  31. Ze Yu Zhang, Arun Verma, Finale Doshi-Velez, and Bryan Kian Hsiang Low. 2025b. Understanding the relationship between prompts and response uncertainty in large language models. Preprint, arXiv:2407.14845. — [arXiv:2407.14845](https://arxiv.org/abs/2407.14845) find
  32. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. Preprint, arXiv:2306.05685.
  33. Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhang-2026-adaptive-prompt-optimization-open-ended/