Elamur Research⌕
☰
home / papers / 2026 / Simple Role Assignment is Extraordinarily Effective for Safety Alignment

Simple Role Assignment is Extraordinarily Effective for Safety Alignment

Paper · 2026 · Zhou Ziheng, Jiakun Ding, Zhaowei Zhang et al.

Summary

The paper addresses a weakness of principle-based safety alignment for LLMs: fixed rule lists can never cover every situation and give little guidance on how a value applies in a specific context. Drawing on Theory of Mind and social role theory, the authors argue that a social role such as "mother" or "judge" carries both the values to uphold and the reasoning schema needed to apply them. They support this with a formal analysis showing that, under idealized assumptions, role conditioning can outperform fixed principle lists. The method is training-free and works at test time. A generator receives a short role prompt, and a few role-based critics then accept or revise its answer over several rounds. Across five model families and several safety benchmarks, the method beats principle-based, CoT and hybrid baselines while leaving general reasoning ability intact. The approach also extends to agentic safety. This matters because it offers a cheap, interpretable alternative for alignment and for building LLM-as-a-Judge evaluators.

Key points

  • Setup and results:
    • Models tested: Qwen3-8B, Gemma3-12B-IT, DeepSeek-V3, Gemini-2.5-Flash and Qwen3-235B.
    • The main configuration uses two "guardianship" roles, "mother" and "principal".
    • With DeepSeek-V3, unsafe outputs on WildJailbreak fell from 81.4% to 3.6%.
    • In an AI-blackmail agent test, unsafe behavior dropped from 65% to 8% with role conditioning alone.
  • Ablations:
    • Concrete roles work better than abstract ones ("mother" beats "parent").
    • Critic feedback is important, with the bulk of the improvement arriving after just one to three rounds of refinement.
    • Including extra roles yields only small further gains.
    • Combining roles with principle-based or CoT prompting outperforms either approach alone.
  • Dynamic variants on SafeEdit:
    • Rewriting the role description for each query added +3.5% (Qwen3-8B) and +5.5% (DeepSeek-V3).
    • Letting the model pick its own role did not beat the fixed "Mother" role.
  • Value landscape: A PCA across nine safety dimensions shows "Mother" is strongest on personal-harm categories, while "Principal" is strong on both personal and public ones, which explains why the pair works well together.
  • Robustness:
    • On a Chinese translation of SafeEdit, the average defense rate was 82.7%, compared with 77.1% in English.
    • Overtly malicious edits to role descriptions were always detected.
    • Only 3 of 29 subtle injections went unflagged.
    • The method is limited by the strength of the base model.

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

🎓Peking UniversityBeijing · China315 works · 43,268 citations🎓Zhejiang UniversityHangzhou · China224 works · 13,587 citations🎓University of California, Los AngelesLos Angeles · United States99 works · 56,330 citations🎓Tianjin UniversityTianjin · China52 works · 1,128 citations🏛️Beijing Institute for General Artificial IntelligenceBeijing · China11 works · 633 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 Zhou Ziheng, Jiakun Ding, Zhaowei Zhang, Ruosen Gao, Ying Nian Wu, Demetri Terzopoulos, Yipeng Kang, Fangwei Zhong and Junqi Wang. 2026. Simple Role Assignment is Extraordinarily Effective for Safety Alignment. 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.1164

@inproceedings{ziheng2026simple,
  title = {{Simple Role Assignment is Extraordinarily Effective for Safety Alignment}},
  author = {Zhou Ziheng, Jiakun Ding, Zhaowei Zhang, Ruosen Gao, Ying Nian Wu, Demetri Terzopoulos, Yipeng Kang, Fangwei Zhong and Junqi Wang},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1164},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1164}
}

References

53 references, 19 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, and 1 others. 2022. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073.
  2. Rewina Bedemariam, Natalie Perez, Sreyoshi Bhaduri, Satya Kapoor, Alex Gil, Elizabeth Conjar, Ikkei Itoku, David Theil, Aman Chadha, and Naumaan Nayyar. 2025. Potential and perils of large language models as judges of unstructured textual data. arXiv preprint arXiv:2501.08167. — [arXiv:2501.08167](https://arxiv.org/abs/2501.08167) find
  3. Rishabh Bhardwaj and Soujanya Poria. 2023. Red-teaming large language models using chain of utterances for safety-alignment. arXiv preprint arXiv:2308.09662.
  4. Bruce J Biddle. 1986. Recent developments in role theory. Annual Review of Sociology, 12:67–92. find
  5. Peter J Burke and Jan E Stets. 2009. Identity Theory. Oxford University Press. find
  6. Hongyu Chen and Seraphina Goldfarb-Tarrant. 2025. Safer or luckier? llms as safety evaluators are not robust to artifacts. arXiv preprint arXiv:2503.09347.
  7. Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. 2017. Deep reinforcement learning from human preferences. Advances in neural information processing systems, 30. find
  8. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, and 1 others. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arX…
  9. Yann Dubois, Balázs Galambosi, Percy Liang, and Tatsunori B Hashimoto. 2024. Length-controlled alpacaeval: A simple way to debias automatic evaluators. arXiv preprint arXiv:2404.04475. find
  10. Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. 2024. Kto: Model alignment as prospect theoretic optimization. arXiv preprint arXiv:2402.01306. find
  11. Chris Frith and Uta Frith. 2005. Theory of mind. Current biology, 15(17):R644–R645. find
  12. Songyang Gao, Qiming Ge, Wei Shen, Shihan Dou, Junjie Ye, Xiao Wang, Rui Zheng, Yicheng Zou, Zhi Chen, Hang Yan, and 1 others. 2024. Linear alignment: A closed-form solution for aligning human preferences without tuning and feedback. arXiv preprint arXiv:2401.11458. — [arXiv:2401.11458](https://arxiv.org/abs/2401.11458) find
  13. Zhiguang Han and Zijian Wang. 2024. Rethinking the role-play prompting in mathematical reasoning tasks. In Proceedings of the 1st Workshop on Efficiency, Security, and Generalization of Multimedia Foundation Models, pages 13–17. find
  14. Dulhan Jayalath, Shashwat Goel, Thomas Foster, Parag Jain, Suchin Gururangan, Cheng Zhang, Anirudh Goyal, and Alan Schelten. 2025. Compute as teacher: Turning inference compute into reference-free supervision. arXiv preprint arXiv:2509.14234. — [arXiv:2509.14234](https://arxiv.org/abs/2509.14234) find
  15. Jiaming Ji, Tianyi Qiu, Boyuan Chen, Borong Zhang, Hantao Lou, Kaile Wang, Yawen Duan, Zhonghao He, Jiayi Zhou, Zhaowei Zhang, and 1 others. 2023. Ai alignment: A comprehensive survey. arXiv preprint arXiv:2310.19852.
  16. Hongchao Jiang, Yiming Chen, Yushi Cao, Hung-yi Lee, and Robby T Tan. 2025. Codejudgebench: Benchmarking llm-as-a-judge for coding tasks. arXiv preprint arXiv:2507.10535.
  17. Liwei Jiang, Kavel Rao, Seungju Han, Allyson Ettinger, Faeze Brahman, Sachin Kumar, Niloofar Mireshghallah, Ximing Lu, Maarten Sap, Yejin Choi, and 1 others. 2024. Wildteaming at scale: From in-the-wild jailbreaks to (adversarially) safer language models. Advances in Neural Information Processing Sy…
  18. Immanuel Kant. 1785. Groundwork of the Metaphysics of Morals. Translated by M. Gregor, Cambridge University Press, 1998. find
  19. Aobo Kong, Shiwan Zhao, Hao Chen, Qicheng Li, Yong Qin, Ruiqi Sun, Xin Zhou, Enzhi Wang, and Xiaohang Dong. 2023. Better zero-shot reasoning with role-play prompting. arXiv preprint arXiv:2308.07702.
  20. Harrison Lee, Samrat Phatale, Hassan Mansoor, Kellie Ren Lu, Thomas Mesnard, Johan Ferret, Colton Bishop, Ethan Hall, Victor Carbune, and Abhinav Rastogi. 2023. Rlaif: Scaling reinforcement learning from human feedback with ai feedback. find
  21. Aidar Myrzakhan, Sondos Mahmoud Bsharat, and Zhiqiang Shen. 2024. Open-llm-leaderboard: From multi-choice to open-style questions for llms evaluation, benchmark, and arena. arXiv preprint arXiv:2406.07545. — [arXiv:2406.07545](https://arxiv.org/abs/2406.07545) find
  22. Sangkyu Lee, Sungdong Kim, Ashkan Yousefpour, Minjoon Seo, Kang Min Yoo, and Youngjae Yu. 2024. Aligning large language models by on-policy self-judgment. arXiv preprint arXiv:2402.11253. find
  23. Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, and 1 others. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730–27744.
  24. 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.
  25. Yuhui Li, Fangyun Wei, Jinjing Zhao, Chao Zhang, and Hongyang Zhang. 2023. Rain: Your language models can align themselves without finetuning. arXiv preprint arXiv:2309.07124. — [arXiv:2309.07124](https://arxiv.org/abs/2309.07124) find
  26. Bill Yuchen Lin, Abhilasha Ravichander, Ximing Lu, Nouha Dziri, Melanie Sclar, Khyathi Chandu, Chandra Bhagavatula, and Yejin Choi. 2023. Urial: Tuning-free instruction learning and alignment for untuned llms. In NeurIPS 2023 Workshop on Instruction Tuning and Instruction Following. find
  27. Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, and 1 others. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437. find
  28. Xianghe Pang, Shuo Tang, Rui Ye, Yuxin Xiong, Bolun Zhang, Yanfeng Wang, and Siheng Chen. 2024. Self-alignment of large language models via monopolylogue-based social scene simulation. arXiv preprint arXiv:2402.05699. — [arXiv:2402.05699](https://arxiv.org/abs/2402.05699) find
  29. Chen Qian, Zihao Xie, Yifei Wang, Wei Liu, Kunlun Zhu, Hanchen Xia, Yufan Dang, Zhuoyun Du, Weize Chen, Cheng Yang, and 1 others. 2024. Scaling large language model-based multi-agent collaboration. arXiv preprint arXiv:2406.07155.
  30. Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. Advances in neural information processing systems, 36:53728–53741.
  31. Yunfan Shao, Linyang Li, Junqi Dai, and Xipeng Qiu. 2023. Character-llm: A trainable agent for role-playing. arXiv preprint arXiv:2310.10158.
  32. Ryan Louie, Ananjan Nandi, William Fang, Cheng Chang, Emma Brunskill, and Diyi Yang. 2024. Roleplay-doh: Enabling domain-experts to create llm-simulated patients via eliciting and adhering to principles. arXiv preprint arXiv:2407.00870.
  33. Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, and 1 others. 2025. Gemma 3 technical report. arXiv preprint arXiv:2503.19786. find
  34. Li-Chun Lu, Shou-Jen Chen, Tsung-Min Pai, Chan-Hung Yu, Hung-yi Lee, and Shao-Hua Sun. 2024. Llm discussion: Enhancing the creativity of large language models via discussion framework and role-play. arXiv preprint arXiv:2405.06373. — [arXiv:2405.06373](https://arxiv.org/abs/2405.06373) find
  35. Anyi Wang, Dong Shu, Yifan Wang, Yunpu Ma, and Mengnan Du. 2025a. Improving llm reasoning through interpretable role-playing steering. arXiv preprint arXiv:2506.07335. find
  36. Kaifeng Lyu, Haoyu Zhao, Xinran Gu, Dingli Yu, Anirudh Goyal, and Sanjeev Arora. 2024. Keeping llms aligned after fine-tuning: The crucial role of prompt templates. Advances in Neural Information Processing Systems, 37:118603–118631. find
  37. Yu Meng, Mengzhou Xia, and Danqi Chen. 2024. Simpo: Simple preference optimization with a reference-free reward. Advances in Neural Information Processing Systems, 37:124198–124235. find
  38. John Stuart Mill. 1863. Utilitarianism. Parker, Son, and Bourn. find
  39. Seumas Miller. 2003. Social institutions. In Realism in action: Essays in the philosophy of the social sciences, pages 233–249. Springer. find
  40. Mengru Wang, Ningyu Zhang, Ziwen Xu, Zekun Xi, Shumin Deng, Yunzhi Yao, Qishen Zhang, Linyi Yang, Jindong Wang, and Huajun Chen. 2024. Detoxifying large language models via knowledge editing. arXiv preprint arXiv:2403.14472.
  41. Yutong Wang, Pengliang Ji, Chaoqun Yang, Kaixin Li, Ming Hu, Jiaoyang Li, and Guillaume Sartoretti. 2025b. Mcts-judge: Test-time scaling in llm-as-a-judge for code correctness evaluation. arXiv preprint arXiv:2502.12468. — [arXiv:2502.12468](https://arxiv.org/abs/2502.12468) find
  42. Zekun Moore Wang, Zhongyuan Peng, Haoran Que, Jiaheng Liu, Wangchunshu Zhou, Yuhan Wu, Hongcheng Guo, Ruitong Gan, Zehao Ni, Jian Yang, and 1 others. 2023. Rolellm: Benchmarking, eliciting, and enhancing role-playing abilities of large language models. arXiv preprint arXiv:2310.00746.
  43. Jiaxin Wen, Zachary Ankner, Arushi Somani, Peter Hase, Samuel Marks, Jacob Goldman-Wetzler, Linda Petrini, Henry Sleight, Collin Burns, He He, and 1 others. 2025. Unsupervised elicitation of language models. arXiv preprint arXiv:2506.10139. find
  44. Tianhao Wu, Weizhe Yuan, Olga Golovneva, Jing Xu, Yuandong Tian, Jiantao Jiao, Jason Weston, and Sainbayar Sukhbaatar. 2024. Meta-rewarding language models: Self-improving alignment with llm-as-ameta-judge. arXiv preprint arXiv:2407.19594. — [arXiv:2407.19594](https://arxiv.org/abs/2407.19594) find
  45. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388.
  46. Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Sainbayar Sukhbaatar, Jing Xu, and Jason Weston. 2024. Self-rewarding language models. arXiv preprint arXiv:2401.10020, 3. find
  47. Wenjun Zeng, Yuchi Liu, Ryan Mullins, Ludovic Peran, Joe Fernandez, Hamza Harkous, Karthik Narasimhan, Drew Proud, Piyush Kumar, Bhaktipriya Radharapu, and 1 others. 2024. Shieldgemma: Generative ai content moderation based on gemma. arXiv preprint arXiv:2407.21772. — [arXiv:2407.21772](https://arxiv.org/abs/2407.21772) find
  48. Xiaoying Zhang, Baolin Peng, Ye Tian, Jingyan Zhou, Lifeng Jin, Linfeng Song, Haitao Mi, and Helen Meng. 2024. Self-alignment for factuality: Mitigating hallucinations in llms via self-evaluation. arXiv preprint arXiv:2402.09267.
  49. Zhaowei Zhang, Fengshuo Bai, Qizhi Chen, Chengdong Ma, Mingzhi Wang, Haoran Sun, Zilong Zheng, and Yaodong Yang. 2025. Amulet: Realignment during test time for personalized preference adaptation of llms. arXiv preprint arXiv:2502.19148. — [arXiv:2502.19148](https://arxiv.org/abs/2502.19148) find
  50. Yuwei Zhao, Ziyang Luo, Yuchen Tian, Hongzhan Lin, Weixiang Yan, Annan Li, and Jing Ma. 2024. Codejudge-eval: Can large language models be good judges in code understanding? arXiv preprint arXiv:2408.10718. find
  51. Xin Zheng, Jie Lou, Boxi Cao, Xueru Wen, Yuqiu Ji, Hongyu Lin, Yaojie Lu, Xianpei Han, Debing Zhang, and Le Sun. 2024. Critic-cot: Boosting the reasoning abilities of large language model via chain-of-thoughts critic. arXiv preprint arXiv:2408.16326. — [arXiv:2408.16326](https://arxiv.org/abs/2408.16326) find
  52. Mingye Zhu, Yi Liu, Lei Zhang, Junbo Guo, and Zhendong Mao. 2025. On-the-fly preference alignment via principle-guided decoding. arXiv preprint arXiv:2502.14204. — [arXiv:2502.14204](https://arxiv.org/abs/2502.14204) find
  53. Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. 2023. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/ziheng-2026-simple-role-assignment-is-extraordinarily/