---
title: "PersonaForge — paper by Tong and Zou (2026)"
description: "LLM-based role-playing agents tend to lose their characters over long conversations: traits shift, speaking style wobbles, and… 18 references."
authors: ["Jizhou Tong", "Sirui Zou"]
year: 2026
venue: "Findings of the Association for Computational Linguistics: ACL 2026"
source: "https://aclanthology.org/2026.findings-acl.386/"
doi: "10.18653/v1/2026.findings-acl.386"
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/tong-2026-personaforge-psychology-grounded-dual-process/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/tong-2026-personaforge-psychology-grounded-dual-process/)

[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/39/) / PersonaForge: Psychology-Grounded Dual-Process Architecture for Personality-Consistent Role-Playing Agents

# PersonaForge: Psychology-Grounded Dual-Process Architecture for Personality-Consistent Role-Playing Agents

**Authors:** Jizhou Tong and Sirui Zou · **Year:** 2026 · **Venue:** [Findings of the Association for Computational Linguistics: ACL 2026](https://md.research.elamur.ai/free/en/venues/findings/) ([Findings of the Association for Computational Linguistics 2026](https://md.research.elamur.ai/free/en/venues/findings/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.findings-acl.386/) · [PDF](https://aclanthology.org/2026.findings-acl.386.pdf) · [DOI 10.18653/v1/2026.findings-acl.386](https://doi.org/10.18653/v1/2026.findings-acl.386) · [OpenAlex](https://openalex.org/W7166818976)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## Summary
LLM-based role-playing agents tend to lose their characters over long conversations: traits shift, speaking style wobbles, and emotional state is ignored. PersonaForge tackles this with a psychology-based design. Each character profile has three layers: core traits (Big Five scores plus one of nine Vaillant defense mechanisms), a speaking-style matrix, and a dynamic state that tracks mood, energy and relationships. Generation follows a dual-process scheme. For emotionally or relationally critical turns, a hidden Inner Monologue is produced first and then rendered in the character's voice. Ordinary turns skip this extra step.

The authors test two claims. The first is that structured psychological dimensions constrain behaviour better than free-text descriptions. The second is that dense constraints need an intermediate cognitive workspace. On 88 annotated characters, the system improves personality consistency, sharply reduces drift over 50 turns, and also generalizes to RoleBench. Profiles can be extracted automatically from raw text, so the method works without fine-tuning. It is therefore useful as a model-agnostic guardrail for consistent character agents.

## Key points
- The dataset covers 88 characters from five literary domains, including classical Chinese literature and Western fantasy. Three experts annotated them, with overall κ = 0.78 (Big Five 0.76, DM 0.82).
- Personality consistency rises 19.4% over a Structured-CoT baseline, and the metric correlates with human judgments at r = 0.82. Drift over 50 turns falls to 6.3%, compared with 24.8% for the baseline. Defense mechanism manifestation increases by 64.7%.
- On RoleBench, the system reaches a 73.2% pairwise win-rate and cuts drift from 20.4% to 8.4%. It stays in character in 95.2% of RoleBreak-style adversarial attacks.
- Selective activation of the dual process fires on roughly 40% of turns. It keeps 96% of full-system performance at 13.4% token overhead.
- Ablations support both claims. Replacing Big Five/DM with natural language lowers PC by 0.18, and removing the dual process drops PC to 0.57, below Structured-CoT's 0.72.
- Main runs use Gemini 2.5 Flash. An open-source DeepSeek-V3 pipeline reaches PC 0.84. Qwen2.5-7B handles only simpler defenses, not "mature" ones such as Humor.

*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

- [Wuhan University](https://md.research.elamur.ai/free/en/universities/wuhan-university/) — University, Wuhan, China · 35 works · 1,127 citations
- [Southwestern University of Finance and Economics](https://md.research.elamur.ai/free/en/universities/southwestern-university-of-finance-and-economics/) — University, Chengdu, China · 4 works · 1 citation

📚 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: Jizhou Tong and Sirui Zou. 2026. PersonaForge: Psychology-Grounded Dual-Process Architecture for Personality-Consistent Role-Playing Agents. Findings of the Association for Computational Linguistics: ACL 2026.

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

```bibtex
@inproceedings{tong2026personaforge,
  title = {{PersonaForge: Psychology-Grounded Dual-Process Architecture for Personality-Consistent Role-Playing Agents}},
  author = {Jizhou Tong and Sirui Zou},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.386},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.386}
}
```

- [← previous PersonaArena: Dynamic Simulation for Evaluating and Enhancing Persona-Level Role-Playing in Large Language Models](https://md.research.elamur.ai/free/en/papers/shi-2026-personaarena-dynamic-simulation-evaluating-enhancing/)
- [next → Personalizing LLMs with Binary Feedback: A Preference-Calibrated Optimization Framework](https://md.research.elamur.ai/free/en/papers/ma-2026-personalizing-llms-binary-feedback-preference/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/39/)

## References

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

1. Roy F Baumeister, Ellen Bratslavsky, Catrin Finkenauer, and Kathleen D Vohs. 2001. Bad is stronger than good. Review of general psychology, 5(4):323– 370. [find](https://scholar.google.com/scholar?q=Roy+F+Baumeister%2C+Ellen+Bratslavsky%2C+Catrin+Finkenauer%2C+and+Kathleen+D+Vohs%2E+2001%2E+Bad+is+stronger+than+good%2E+Review+of+general+psychology%2C+5%284%29%3A323%E2%80%93+370%2E)
2. Jiangjie Chen, Xintao Wang, Rui Xu, Siyu Yuan, Yikai Zhang, Wei Shi, Jian Xie, Shuang Li, Ruihan Yang, Tinghui Zhu, and 1 others. 2024. [From persona to personalization: A survey on role-playing language agents](https://md.research.elamur.ai/free/en/papers/chen-2024-persona-personalization-survey-role-playing/). arXiv preprint arXiv:2404.18231.
3. Xiachong Feng, Longxu Dou, and Lingpeng Kong. 2025. Reasoning does not necessarily improve role-playing ability. arXiv preprint arXiv:2502.16940. [find](https://scholar.google.com/scholar?q=Xiachong+Feng%2C+Longxu+Dou%2C+and+Lingpeng+Kong%2E+2025%2E+Reasoning+does+not+necessarily+improve+role%2Dplaying+ability%2E+arXiv+preprint+arXiv%3A2502%2E16940%2E)
4. Daniel Kahneman. 2011. macmillan. [find](https://scholar.google.com/scholar?q=Daniel+Kahneman%2E+2011%2E+macmillan%2E)
5. Thinking, fast and slow. Charles Packer, Vivian Fang, Shishir_G Patil, Kevin Lin, Sarah Wooders, and Joseph_E Gonzalez. 2023. Memgpt: Towards llms as operating systems. [find](https://scholar.google.com/scholar?q=Thinking%2C+fast+and+slow%2E+Charles+Packer%2C+Vivian+Fang%2C+Shishir%5FG+Patil%2C+Kevin+Lin%2C+Sarah+Wooders%2C+and+Joseph%5FE+Gonzalez%2E+2023%2E+Memgpt%3A+Towards+llms+as+operating+systems%2E)
6. Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology, pages 1–22. [find](https://scholar.google.com/scholar?q=Joon+Sung+Park%2C+Joseph+O%E2%80%99Brien%2C+Carrie+Jun+Cai%2C+Meredith+Ringel+Morris%2C+Percy+Liang%2C+and+Michael+S+Bernstein%2E+2023%2E+Generative+agents%3A+Interactive+simulacra+of+human+behavior%2E+In+Proceedings+of+the+36)
7. Yiting Ran, Xintao Wang, Tian Qiu, Jiaqing Liang, Yanghua Xiao, and Deqing Yang. 2025. Bookworld: From novels to interactive agent societies for creative story generation. arXiv preprint arXiv:2504.14538. — [arXiv:2504.14538](https://arxiv.org/abs/2504.14538) [find](https://scholar.google.com/scholar?q=Yiting+Ran%2C+Xintao+Wang%2C+Tian+Qiu%2C+Jiaqing+Liang%2C+Yanghua+Xiao%2C+and+Deqing+Yang%2E+2025%2E+Bookworld%3A+From+novels+to+interactive+agent+societies+for+creative+story+generation%2E+arXiv+preprint+arXiv%3A2504%2E14)
8. Yunfan Shao, Linyang Li, Junqi Dai, and Xipeng Qiu. 2023. [Character-llm: A trainable agent for role-playing](https://md.research.elamur.ai/free/en/papers/shao-2023-character-llm-trainable-agent-role/). arXiv preprint arXiv:2310.10158.
9. Zhouxing Tan, Ruochong Xiong, Yulong Wan, Jinlong Ma, Hanlin Xue, Qichun Deng, Haifeng Jing, Zhengtong Zhang, Depei Liu, Shiyuan Luo, and 1 others. 2025. Detecting emotional dynamic trajectories: An evaluation framework for emotional support in language models. arXiv preprint arXiv:2511.09003. — [arXiv:2511.09003](https://arxiv.org/abs/2511.09003) [find](https://scholar.google.com/scholar?q=Zhouxing+Tan%2C+Ruochong+Xiong%2C+Yulong+Wan%2C+Jinlong+Ma%2C+Hanlin+Xue%2C+Qichun+Deng%2C+Haifeng+Jing%2C+Zhengtong+Zhang%2C+Depei+Liu%2C+Shiyuan+Luo%2C+and+1+others%2E+2025%2E+Detecting+emotional+dynamic+trajectories%3A+An+e)
10. Yihong Tang, Kehai Chen, Muyun Yang, Zhengyu Niu, Jing Li, Tiejun Zhao, and Min Zhang. 2025. Thinking in character: Advancing role-playing agents with role-aware reasoning. arXiv preprint arXiv:2506.01748. — [arXiv:2506.01748](https://arxiv.org/abs/2506.01748) [find](https://scholar.google.com/scholar?q=Yihong+Tang%2C+Kehai+Chen%2C+Muyun+Yang%2C+Zhengyu+Niu%2C+Jing+Li%2C+Tiejun+Zhao%2C+and+Min+Zhang%2E+2025%2E+Thinking+in+character%3A+Advancing+role%2Dplaying+agents+with+role%2Daware+reasoning%2E+arXiv+preprint+arXiv%3A2506%2E0)
11. George E Vaillant. 1994. Ego mechanisms of defense and personality psychopathology. Journal of abnormal psychology, 103(1):44. [find](https://scholar.google.com/scholar?q=George+E+Vaillant%2E+1994%2E+Ego+mechanisms+of+defense+and+personality+psychopathology%2E+Journal+of+abnormal+psychology%2C+103%281%29%3A44%2E)
12. Lei Wang, Jianxun Lian, Yi Huang, Yanqi Dai, Haoxuan Li, Xu Chen, Xing Xie, and Ji-Rong Wen. 2025. Characterbox: Evaluating the role-playing capabilities of llms in text-based virtual worlds. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computat… [find](https://scholar.google.com/scholar?q=Lei+Wang%2C+Jianxun+Lian%2C+Yi+Huang%2C+Yanqi+Dai%2C+Haoxuan+Li%2C+Xu+Chen%2C+Xing+Xie%2C+and+Ji%2DRong+Wen%2E+2025%2E+Characterbox%3A+Evaluating+the+role%2Dplaying+capabilities+of+llms+in+text%2Dbased+virtual+worlds%2E+In+Proce)
13. Noah Wang, Zy Peng, Haoran Que, Jiaheng Liu, Wangchunshu Zhou, Yuhan Wu, Hongcheng Guo, Ruitong Gan, Zehao Ni, Jian Yang, and 1 others. 2024a. [Rolellm: Benchmarking, eliciting, and enhancing role-playing abilities of large language models](https://md.research.elamur.ai/free/en/papers/wang-2023-rolellm-benchmarking-eliciting-enhancing-role/). In Findings of the Association for Computational Linguistics…
14. Xintao Wang, Yunze Xiao, Jen-tse Huang, Siyu Yuan, Rui Xu, Haoran Guo, Quan Tu, Yaying Fei, Ziang Leng, Wei Wang, and 1 others. 2024b. [Incharacter: Evaluating personality fidelity in role-playing agents through psychological interviews](https://md.research.elamur.ai/free/en/papers/wang-2024-incharacter-evaluating-personality-fidelity-role/). In Proceedings of the 62nd Annual Meeting of the Association fo…
15. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. [Chain-of-thought prompting elicits reasoning in large language models](https://md.research.elamur.ai/free/en/papers/wei-2022-chain-thought-prompting-elicits-reasoning/). Advances in neural information processing systems, 35:24824–24837.
16. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. 2022. [React: Synergizing reasoning and acting in language models](https://md.research.elamur.ai/free/en/papers/yao-2022-react-synergizing-reasoning-acting-language/). In The eleventh international conference on learning representations.
17. Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. 2018. [Personalizing dialogue agents: I have a dog, do you have pets too?](https://md.research.elamur.ai/free/en/papers/zhang-2018-personalizing-dialogue-agents-i-have/) arXiv preprint arXiv:1801.07243.
18. Peixiang Zhong, Chen Zhang, Hao Wang, Yong Liu, and Chunyan Miao. 2020. [Towards persona-based empathetic conversational models](https://md.research.elamur.ai/free/en/papers/zhong-2020-persona-based-empathetic-conversational-models/). arXiv preprint arXiv:2004.12316.

*Updated 1 November 2026*