---
title: "Persona Prompting as a Lens on LLM Social Reasoning (2026)"
description: "They run three LLMs and compare outputs against neutral, persona-free baselines. 24 references."
authors: ["Jing Yang", "Moritz Hechtbauer", "Elisabeth Khalilov", "Evelyn Luise Brinkmann", "Vera Schmitt", "Nils Feldhus"]
year: 2026
venue: "Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)"
source: "https://aclanthology.org/2026.eacl-long.52/"
doi: "10.18653/v1/2026.eacl-long.52"
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/yang-2026-persona-prompting-as-lens-llm/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/yang-2026-persona-prompting-as-lens-llm/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/8/) / Persona Prompting as a Lens on LLM Social Reasoning

# Persona Prompting as a Lens on LLM Social Reasoning

**Authors:** [Jing Yang](https://md.research.elamur.ai/free/en/authors/jing-yang/), Moritz Hechtbauer, Elisabeth Khalilov, Evelyn Luise Brinkmann, [Vera Schmitt](https://md.research.elamur.ai/free/en/authors/vera-schmitt/) and [Nils Feldhus](https://md.research.elamur.ai/free/en/authors/nils-feldhus/) · **Year:** 2026 · **Venue:** [Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)](https://md.research.elamur.ai/free/en/venues/eacl/) ([EACL 2026](https://md.research.elamur.ai/free/en/venues/eacl/2026/)) · **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/2026.eacl-long.52/) · [PDF](https://aclanthology.org/2026.eacl-long.52.pdf) · [DOI 10.18653/v1/2026.eacl-long.52](https://doi.org/10.18653/v1/2026.eacl-long.52) · [OpenAlex](https://openalex.org/W7140092936)

**Citations:** 2 citations (OpenAlex, 2026-10-14)

## Summary
The paper asks whether persona prompting (PP), where an LLM is told to answer as someone with given demographic traits, changes the word-level rationales it gives along with its labels, and whether those simulated viewpoints match real human annotators. The authors pair three tasks of differing subjectivity (hate speech detection on HateXplain, plus commonsense reasoning and sentiment analysis from the BRWRR re-annotations of CoS-E and SST-2) with single-attribute and composite demographic personas. They run three LLMs and compare outputs against neutral, persona-free baselines. Label quality and rationale overlap are measured against ground truth and against annotations from specific demographic groups, and agreement across personas is also checked. PP helps label prediction on hate speech but tends to lower rationale quality there. Personas do not match their real-world demographic counterparts, and models barely shift under different personas. Biases toward certain annotator groups and a habit of over-flagging harmful content persist regardless of persona. The work suggests caution about using PP for personalization or as a population simulator.

## Key points
- **Personas:** There are 21 single-attribute personas, adapted from the 33 attributes of Castricato et al. (2025) and grouped under seven dimensions: political view, religion, race, education, gender, age and loneliness. Another 12 composite personas combine age, gender and ethnicity so they can be matched to BRWRR's six annotator groups.
- **Data:** The HateXplain subset has 500 posts (301 Normal, 75 Offensive, 124 Hate speech). The BRWRR subsets are CoS-E (500 instances) and SST-2 (263 instances).
- **Models and setup:** Mistral-Medium, Qwen3-32B and GPT-OSS-120B were each run three times. Prompts used in-character step-by-step reasoning and JSON output; removing the reasoning step hurt results, especially rationale selection.
- **Metrics:** Accuracy, Macro-F1, MAE and Mean Error cover labels. Token-F1 and IOU-F1 cover rationales. Krippendorff's α measures inter-persona agreement, which was high for all models except Qwen3.
- **Bias findings:** Older Black and White annotator groups (BO/WO) are matched better than the other groups under every prompting condition. On HateXplain, positive Mean Error shows consistent over-flagging, with Normal posts the hardest class.

*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

- [Technische Universität Berlin](https://md.research.elamur.ai/free/en/universities/technische-universitat-berlin/) — University, Berlin, Germany · 22 works · 957 citations

Also on the byline: [Berlin Institute for the Foundations of Learning and Data](https://ror.org/05dsfb086).

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 17,450 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 17,296 works

🧭 Topics (2): [🧠 Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) · 3,894 works · [🧮 Reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/) · 1,685 works

## Cite

Cite as: Jing Yang, Moritz Hechtbauer, Elisabeth Khalilov, Evelyn Luise Brinkmann, Vera Schmitt and Nils Feldhus. 2026. Persona Prompting as a Lens on LLM Social Reasoning. Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers).

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

```bibtex
@inproceedings{yang2026persona,
  title = {{Persona Prompting as a Lens on LLM Social Reasoning}},
  author = {Jing Yang, Moritz Hechtbauer, Elisabeth Khalilov, Evelyn Luise Brinkmann, Vera Schmitt and Nils Feldhus},
  booktitle = {Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.eacl-long.52},
  url = {https://doi.org/10.18653/v1/2026.eacl-long.52}
}
```

- [← previous Persona Jailbreaking in Large Language Models](https://md.research.elamur.ai/free/en/papers/sandhan-2026-persona-jailbreaking-large-language-models/)
- [next → PersonalAlign: Hierarchical Implicit Intent Alignment for Personalized GUI Agent with Long-Term User-Centric Records](https://md.research.elamur.ai/free/en/papers/lyu-2026-personalalign-hierarchical-implicit-intent-alignment/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/8/)

## References

24 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Lisa P. Argyle, Ethan C. Busby, Nancy Fulda, Joshua R. Gubler, Christopher Rytting, and David Wingate. 2023. [Out of one, many: Using language models to simulate human samples](https://md.research.elamur.ai/free/en/papers/argyle-2022-out-one-many-using-language/). Political Analysis, 31(3):337–351.
2. Louis Castricato, Nathan Lile, Rafael Rafailov, Jan-Philipp Fränken, and Chelsea Finn. 2025. PER-SONA: A reproducible testbed for pluralistic alignment. In Proceedings of the 31st International Conference on Computational Linguistics, pages 11348– 11368, Abu Dhabi, UAE. Association for Computational… [find](https://scholar.google.com/scholar?q=Louis+Castricato%2C+Nathan+Lile%2C+Rafael+Rafailov%2C+Jan%2DPhilipp+Fr%C3%A4nken%2C+and+Chelsea+Finn%2E+2025%2E+PER%2DSONA%3A+A+reproducible+testbed+for+pluralistic+alignment%2E+In+Proceedings+of+the+31st+International+Confer)
3. Myra Cheng, Esin Durmus, and Dan Jurafsky. 2023. [Marked personas: Using natural language prompts to measure stereotypes in language models](https://md.research.elamur.ai/free/en/papers/cheng-2023-marked-personas-natural-language-prompts/). In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1504–1532, Toronto, Canada. Associatio…
4. Ameet Deshpande, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, and Karthik Narasimhan. 2023. [Toxicity in chatgpt: Analyzing persona-assigned language models](https://md.research.elamur.ai/free/en/papers/deshpande-2023-toxicity-chatgpt-analyzing-persona-assigned/). In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 1236–1270, Singapore. Association for Computational Ling…
5. Torrence S Farmer and Chien-Ju Ho. 2025. Examining interactions between user characteristics and explanation modalities on inducing complementarity. In Proceedings of the Extended Abstracts of the CHI Conference on Human Factors in Computing Systems, CHI EA ’25, New York, NY, USA. Association for Co… [find](https://scholar.google.com/scholar?q=Torrence+S+Farmer+and+Chien%2DJu+Ho%2E+2025%2E+Examining+interactions+between+user+characteristics+and+explanation+modalities+on+inducing+complementarity%2E+In+Proceedings+of+the+Extended+Abstracts+of+the+CHI)
6. Shashank Gupta, Vaishnavi Shrivastava, Ameet Deshpande, Ashwin Kalyan, Peter Clark, Ashish Sabharwal, and Tushar Khot. 2024. [Bias runs deep: Implicit reasoning biases in persona-assigned LLMs](https://md.research.elamur.ai/free/en/papers/gupta-2023-bias-runs-deep-implicit-reasoning/). In The Twelfth International Conference on Learning Representations.
7. Tiancheng Hu and Nigel Collier. 2024. [Quantifying the persona effect in LLM simulations](https://md.research.elamur.ai/free/en/papers/hu-2024-quantifying-persona-effect-llm-simulations/). In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10289–10307, Bangkok, Thailand. Association for Computational Linguistics.
8. Jakub Jeck, Florian Leiser, Anne Hüsges, and Ali Sunyaev. 2025. Tell-me: Toward personalized explanations of large language models. In Proceedings of the Extended Abstracts of the CHI Conference on Human Factors in Computing Systems, CHI EA ’25, New York, NY, USA. Association for Computing Machinery… [find](https://scholar.google.com/scholar?q=Jakub+Jeck%2C+Florian+Leiser%2C+Anne+H%C3%BCsges%2C+and+Ali+Sunyaev%2E+2025%2E+Tell%2Dme%3A+Toward+personalized+explanations+of+large+language+models%2E+In+Proceedings+of+the+Extended+Abstracts+of+the+CHI+Conference+on+Hu)
9. Brihi Joshi, Xiang Ren, Swabha Swayamdipta, Rik Koncel-Kedziorski, and Tim Paek. 2025. Improving language model personas via rationalization with psychological scaffolds. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 21747–21770, Suzhou, China. Association for Compu… [find](https://scholar.google.com/scholar?q=Brihi+Joshi%2C+Xiang+Ren%2C+Swabha+Swayamdipta%2C+Rik+Koncel%2DKedziorski%2C+and+Tim+Paek%2E+2025%2E+Improving+language+model+personas+via+rationalization+with+psychological+scaffolds%2E+In+Findings+of+the+Associatio)
10. Malik Khadar, Amoligha Timma, Luka Ludden, and Harmanpreet Kaur. 2025. Explain it like i’m me: User characteristics in xai decision-making. In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency, FAccT ’25, page 1729–1744, New York, NY, USA. Association for Computing… [find](https://scholar.google.com/scholar?q=Malik+Khadar%2C+Amoligha+Timma%2C+Luka+Ludden%2C+and+Harmanpreet+Kaur%2E+2025%2E+Explain+it+like+i%E2%80%99m+me%3A+User+characteristics+in+xai+decision%2Dmaking%2E+In+Proceedings+of+the+2025+ACM+Conference+on+Fairness%2C+Accou)
11. Junseok Kim, Nakyeong Yang, and Kyomin Jung. 2024. Persona is a double-edged sword: Mitigating the negative impact of role-playing prompts in zero-shot reasoning tasks. arXiv, abs/2408.08631. — [arXiv:2408.08631](https://arxiv.org/abs/2408.08631) [find](https://scholar.google.com/scholar?q=Junseok+Kim%2C+Nakyeong+Yang%2C+and+Kyomin+Jung%2E+2024%2E+Persona+is+a+double%2Dedged+sword%3A+Mitigating+the+negative+impact+of+role%2Dplaying+prompts+in+zero%2Dshot+reasoning+tasks%2E+arXiv%2C+abs%2F2408%2E08631%2E+%E2%80%94+%5BarXiv)
12. Hannah Rose Kirk, Bertie Vidgen, Paul Röttger, and Scott A Hale. 2024. The benefits, risks and bounds of personalizing the alignment of large language models to individuals. Nature Machine Intelligence, 6(4):383–392. [find](https://scholar.google.com/scholar?q=Hannah+Rose+Kirk%2C+Bertie+Vidgen%2C+Paul+R%C3%B6ttger%2C+and+Scott+A+Hale%2E+2024%2E+The+benefits%2C+risks+and+bounds+of+personalizing+the+alignment+of+large+language+models+to+individuals%2E+Nature+Machine+Intelligenc)
13. Andy Liu, Mona Diab, and Daniel Fried. 2024. Evaluating large language model biases in persona-steered generation. In Findings of the Association for Computational Linguistics: ACL 2024, pages 9832–9850, Bangkok, Thailand. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Andy+Liu%2C+Mona+Diab%2C+and+Daniel+Fried%2E+2024%2E+Evaluating+large+language+model+biases+in+persona%2Dsteered+generation%2E+In+Findings+of+the+Association+for+Computational+Linguistics%3A+ACL+2024%2C+pages+9832%E2%80%9398)
14. Binny Mathew, Punyajoy Saha, Seid Muhie Yimam, Chris Biemann, Pawan Goyal, and Animesh Mukherjee. 2021. Hatexplain: A benchmark dataset for explainable hate speech detection. Proceedings of the AAAI Conference on Artificial Intelligence, 35(17):14867–14875. and biases. In Thirty-seventh Conference o… [find](https://scholar.google.com/scholar?q=Binny+Mathew%2C+Punyajoy+Saha%2C+Seid+Muhie+Yimam%2C+Chris+Biemann%2C+Pawan+Goyal%2C+and+Animesh+Mukherjee%2E+2021%2E+Hatexplain%3A+A+benchmark+dataset+for+explainable+hate+speech+detection%2E+Proceedings+of+the+AAAI+C)
15. Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. [Recursive deep models for semantic compositionality over a sentiment treebank](https://md.research.elamur.ai/free/en/papers/socher-2013-recursive-deep-models-semantic-compositionality/). In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631–…
16. Kacper Sokol and Peter Flach. 2020. One explanation does not fit all. KI - Künstliche Intelligenz, 34(2):235– 250. [find](https://scholar.google.com/scholar?q=Kacper+Sokol+and+Peter+Flach%2E+2020%2E+One+explanation+does+not+fit+all%2E+KI+%2D+K%C3%BCnstliche+Intelligenz%2C+34%282%29%3A235%E2%80%93+250%2E)
17. Pia Sommerauer, Giulia Rambelli, and Tommaso Caselli. 2025. Simulating identity, propagating bias: Abstraction and stereotypes in LLM-generated text. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 19812–19831, Suzhou, China. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Pia+Sommerauer%2C+Giulia+Rambelli%2C+and+Tommaso+Caselli%2E+2025%2E+Simulating+identity%2C+propagating+bias%3A+Abstraction+and+stereotypes+in+LLM%2Dgenerated+text%2E+In+Findings+of+the+Association+for+Computational+L)
18. Terne Sasha Thorn Jakobsen, Laura Cabello, and Anders Søgaard. 2023. Being right for whose right reasons? In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1033–1054, Toronto, Canada. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Terne+Sasha+Thorn+Jakobsen%2C+Laura+Cabello%2C+and+Anders+S%C3%B8gaard%2E+2023%2E+Being+right+for+whose+right+reasons%3F+In+Proceedings+of+the+61st+Annual+Meeting+of+the+Association+for+Computational+Linguistics+%28Vo)
19. Robert Nimmo, Marios Constantinides, Ke Zhou, Daniele Quercia, and Simone Stumpf. 2024. User characteristics in explainable ai: The rabbit hole of personalization? In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, CHI ’24, New York, NY, USA. Association for Computing M… [find](https://scholar.google.com/scholar?q=Robert+Nimmo%2C+Marios+Constantinides%2C+Ke+Zhou%2C+Daniele+Quercia%2C+and+Simone+Stumpf%2E+2024%2E+User+characteristics+in+explainable+ai%3A+The+rabbit+hole+of+personalization%3F+In+Proceedings+of+the+2024+CHI+Confe)
20. Vishakh Padmakumar and He He. 2024. [Does writing with language models reduce content diversity?](https://md.research.elamur.ai/free/en/papers/padmakumar-2023-writing-language-models-reduce-content/) In The Twelfth International Conference on Learning Representations.
21. Pengshuo Qiu, Frank Rudzicz, and Zining Zhu. 2024. Scenarios and approaches for situated natural language explanations. arXiv, abs/2406.05035. — [arXiv:2406.05035](https://arxiv.org/abs/2406.05035) [find](https://scholar.google.com/scholar?q=Pengshuo+Qiu%2C+Frank+Rudzicz%2C+and+Zining+Zhu%2E+2024%2E+Scenarios+and+approaches+for+situated+natural+language+explanations%2E+arXiv%2C+abs%2F2406%2E05035%2E+%E2%80%94+%5BarXiv%3A2406%2E05035%5D%28https%3A%2F%2Farxiv%2Eorg%2Fabs%2F2406%2E05035%29)
22. Nazneen Fatema Rajani, Bryan McCann, Caiming Xiong, and Richard Socher. 2019. [Explain yourself! leveraging language models for commonsense reasoning](https://md.research.elamur.ai/free/en/papers/rajani-2019-explain-yourself-leveraging-language-models/). In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4932–4942, Florence, Italy. Association for Computat…
23. Richard Rogers and Xiaoke Zhang. 2025. A bias towards neutrality? how llm guardrail sensitivity affects classification. Communication and Change, 1(1):1–18. [find](https://scholar.google.com/scholar?q=Richard+Rogers+and+Xiaoke+Zhang%2E+2025%2E+A+bias+towards+neutrality%3F+how+llm+guardrail+sensitivity+affects+classification%2E+Communication+and+Change%2C+1%281%29%3A1%E2%80%9318%2E)
24. Leonard Salewski, Stephan Alaniz, Isabel Rio-Torto, Eric Schulz, and Zeynep Akata. 2023. In-context impersonation reveals large language models’ strengths [find](https://scholar.google.com/scholar?q=Leonard+Salewski%2C+Stephan+Alaniz%2C+Isabel+Rio%2DTorto%2C+Eric+Schulz%2C+and+Zeynep+Akata%2E+2023%2E+In%2Dcontext+impersonation+reveals+large+language+models%E2%80%99+strengths)

*Updated 2 November 2026*