---
title: "Exploring Large Language Models for Communication Games:"
description: "This paper tests whether frozen large language models can play such games without fine-tuning… 34 references, cited by 70 papers here."
authors: ["Yuzhuang Xu", "Shuo Wang", "Peng Li", "Fuwen Luo", "Xiaolong Wang", "Weidong Liu", "Yang Liu"]
year: 2023
venue: "arXiv:2309.04658"
source: "https://arxiv.org/abs/2309.04658"
doi: "10.48550/arxiv.2309.04658"
arxiv: "2309.04658"
licence: "CC BY 4.0 (arXiv)"
licence_url: "http://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/xu-2023-exploring-large-language-models-communication/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/xu-2023-exploring-large-language-models-communication/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2023](https://md.research.elamur.ai/free/en/papers/2023/page/6/) / Exploring Large Language Models for Communication Games: An Empirical Study on Werewolf

# Exploring Large Language Models for Communication Games: An Empirical Study on Werewolf

**Authors:** Yuzhuang Xu, [Shuo Wang](https://md.research.elamur.ai/free/en/authors/shuo-wang/), [Peng Li](https://md.research.elamur.ai/free/en/authors/peng-li/), Fuwen Luo, [Xiaolong Wang](https://md.research.elamur.ai/free/en/authors/xiaolong-wang/), Weidong Liu and [Yang Liu](https://md.research.elamur.ai/free/en/authors/yang-liu/) · **Year:** 2023 · **Venue:** arXiv:2309.04658 · **Licence:** [CC BY 4.0 (arXiv)](http://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/2309.04658) · [PDF](https://arxiv.org/pdf/2309.04658) · [DOI 10.48550/arxiv.2309.04658](https://doi.org/10.48550/arxiv.2309.04658) · [arXiv 2309.04658](https://arxiv.org/abs/2309.04658) · [OpenAlex](https://openalex.org/W4386644471)

**Citations:** 24 citations (OpenAlex, 2026-10-14) · [70](https://md.research.elamur.ai/free/en/papers/xu-2023-exploring-large-language-models-communication/cited-by/) in our library, of the 141,840 papers whose bibliographies we read

## Summary
Communication games are incomplete-information games in which players must reason through natural language. They are hard for AI. Earlier Werewolf agents relied on rules or speech templates, and strong Diplomacy systems needed large amounts of human data. This paper tests whether frozen large language models can play such games without fine-tuning, using Werewolf as the test case. The seven players are each an LLM agent driven by a structured prompt, and the roles consist of a pair of werewolves, a pair of villagers, and one seer, one guard and one witch. To fit limited context windows, each agent gets a compressed history. It contains the latest messages, heuristically selected informative statements, and the agent's own reflection on the situation. The reflection also serves as a reasoning step, much like chain-of-thought. To improve over repeated games, the framework scores earlier response–reflection pairs and retrieves advice that fits the current state. This is non-parametric learning that works across rounds and with many players. Experiments show the agents can play and learn from experience. Unprogrammed strategic behaviours also emerge: trust, confrontation, camouflage and leadership. This makes LLMs promising agents for language-heavy social games.

## Key points
- Game setup: 7 players with 5 roles (2 werewolves, 2 villagers, witch, guard, seer). Each role is played by a separate prompted LLM agent, and the full prompt is given in Appendix A.5.
- The prompt has four parts: (1) rules, role descriptions and basic strategy priors; (2) compressed history; (3) suggestions drawn from past experience; (4) a chain-of-thought instruction.
- History is selected by three criteria: freshness (the most recent K messages), informativeness (heuristically chosen key messages) and completeness (the agent's reflection).
- Experience learning needs no parameter updates or human-annotated data. Unlike Shinn et al. (2023), which uses experience within a single round, and Fu et al. (2023), which handles two-player games, this method uses experience across rounds and supports multiple players.
- Observed emergent behaviours include werewolves posing as villagers, a seer inventing a pretext to signal a werewolf without revealing its role, and calls to action meant to steer votes. The authors argue that most of this is rational play rather than hallucination (Appendix A.4).

*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

- [Tsinghua University](https://md.research.elamur.ai/free/en/universities/tsinghua-university/) — University, Beijing, China · 137 works · 83,262 citations

📚 Shelves (6): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 9,085 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 8,888 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 9,296 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 9,206 works · [🎯 Decision Sciences](https://md.research.elamur.ai/free/en/shelves/decision-sciences/) · 85 works · [🎯 Management Science and Operations Research](https://md.research.elamur.ai/free/en/shelves/decision-sciences/management-science-and-operations-research/) · 36 works

🧭 Topics (1): [🧠 Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) · 1,344 works

## Cite

Cite as: Yuzhuang Xu, Shuo Wang, Peng Li, Fuwen Luo, Xiaolong Wang, Weidong Liu and Yang Liu. 2023. Exploring Large Language Models for Communication Games: An Empirical Study on Werewolf. arXiv:2309.04658.

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

```bibtex
@misc{xu2023exploring,
  title = {{Exploring Large Language Models for Communication Games: An Empirical Study on Werewolf}},
  author = {Yuzhuang Xu, Shuo Wang, Peng Li, Fuwen Luo, Xiaolong Wang, Weidong Liu and Yang Liu},
  year = {2023},
  doi = {10.48550/arxiv.2309.04658},
  eprint = {2309.04658},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2309.04658}
}
```

- [← previous Examining Inter-Consistency of Large Language Models Collaboration: An In-depth Analysis via Debate](https://md.research.elamur.ai/free/en/papers/xiong-2023-examining-inter-consistency-large-language/)
- [next → FLASK: Fine-grained Language Model Evaluation based on Alignment Skill Sets](https://md.research.elamur.ai/free/en/papers/ye-2023-flask-fine-grained-language-model/)
- [↑ back to papers of 2023](https://md.research.elamur.ai/free/en/papers/2023/page/6/)

## References

34 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Michael Bowling, Neil Burch, Michael Johanson, and Oskari Tammelin. 2015. Heads-up limit hold’em poker is solved. Science, 347(6218):145–149. [find](https://scholar.google.com/scholar?q=Michael+Bowling%2C+Neil+Burch%2C+Michael+Johanson%2C+and+Oskari+Tammelin%2E+2015%2E+Heads%2Dup+limit+hold%E2%80%99em+poker+is+solved%2E+Science%2C+347%286218%29%3A145%E2%80%93149%2E)
2. Noam Brown and Tuomas Sandholm. 2018. Superhuman AI for heads-up no-limit poker: Libratus beats top professionals. Science, 359(6374):418–424. [find](https://scholar.google.com/scholar?q=Noam+Brown+and+Tuomas+Sandholm%2E+2018%2E+Superhuman+AI+for+heads%2Dup+no%2Dlimit+poker%3A+Libratus+beats+top+professionals%2E+Science%2C+359%286374%29%3A418%E2%80%93424%2E)
3. Noam Brown and Tuomas Sandholm. 2019. Superhuman AI for multiplayer poker. Science, 365(6456):885–890. [find](https://scholar.google.com/scholar?q=Noam+Brown+and+Tuomas+Sandholm%2E+2019%2E+Superhuman+AI+for+multiplayer+poker%2E+Science%2C+365%286456%29%3A885%E2%80%93890%2E)
4. Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. 2023. [Sparks of artificial general intelligence: Early experiments with GPT-4](https://md.research.elamur.ai/free/en/papers/bubeck-2023-sparks-artificial-general-intelligence-early/). arXiv preprint arXiv:2303.12712.
5. Andrew M Dai and Quoc V Le. 2015. Semi-supervised sequence learning. Advances in NeurIPS 2015, 28. \#\# Limitations FAIR, Anton Bakhtin, Noam Brown, Emily Dinan, Gabriele Farina, Colin Flaherty, Daniel Fried, Andrew Goff, Jonathan Gray, Hengyuan Hu, et al. 2022. Human-level play in the game of diploma… [find](https://scholar.google.com/scholar?q=Andrew+M+Dai+and+Quoc+V+Le%2E+2015%2E+Semi%2Dsupervised+sequence+learning%2E+Advances+in+NeurIPS+2015%2C+28%2E+%5C%23%5C%23+Limitations+FAIR%2C+Anton+Bakhtin%2C+Noam+Brown%2C+Emily+Dinan%2C+Gabriele+Farina%2C+Colin+Flaherty%2C+Danie)
6. Although we have demonstrated that our method possesses the potential to play communication games, there are still some limitations. Firstly, hallucinations (Ji et al., 2023) affect the factuality of the generated content and may negatively impact the reasoning abilities. Then, there may be a larger… [find](https://scholar.google.com/scholar?q=Although+we+have+demonstrated+that+our+method+possesses+the+potential+to+play+communication+games%2C+there+are+still+some+limitations%2E+Firstly%2C+hallucinations+%28Ji+et+al%2E%2C+2023%29+affect+the+factuality+of+)
7. Yao Fu, Hao Peng, Tushar Khot, and Mirella Lapata. 2023. [Improving language model negotiation with self-play and in-context learning from AI feedback](https://md.research.elamur.ai/free/en/papers/fu-2023-improving-language-model-negotiation-self/). arXiv preprint arXiv:2305.10142.
8. Robert Gibbons. 1992. A primer in game theory. [find](https://scholar.google.com/scholar?q=Robert+Gibbons%2E+1992%2E+A+primer+in+game+theory%2E)
9. Yuya Hirata, Michimasa Inaba, Kenichi Takahashi, Fujio Toriumi, Hirotaka Osawa, Daisuke Katagami, and Kousuke Shinoda. 2016. Werewolf game modeling using action probabilities based on play log analysis. In Computers and Games: 9th International Conference, pages 103–114. Springer. \#\# Ethics Statemen… [find](https://scholar.google.com/scholar?q=Yuya+Hirata%2C+Michimasa+Inaba%2C+Kenichi+Takahashi%2C+Fujio+Toriumi%2C+Hirotaka+Osawa%2C+Daisuke+Katagami%2C+and+Kousuke+Shinoda%2E+2016%2E+Werewolf+game+modeling+using+action+probabilities+based+on+play+log+analysi)
10. This study involves the discussion and analysis of a simulated game setting, and any references to “killing”, “eliminating” or related actions are strictly confined within the context of this game. The authors do not condone violence, or illegal activities in any form in real-life scenarios. The gam… [find](https://scholar.google.com/scholar?q=This+study+involves+the+discussion+and+analysis+of+a+simulated+game+setting%2C+and+any+references+to+%E2%80%9Ckilling%E2%80%9D%2C+%E2%80%9Celiminating%E2%80%9D+or+related+actions+are+strictly+confined+within+the+context+of+this+game%2E+Th)
11. Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. [Survey of hallucination in natural language generation](https://md.research.elamur.ai/free/en/papers/ji-2022-survey-hallucination-natural-language-generation/). ACM Computing Surveys, 55(12):1–38.
12. Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. [Large language models are zero-shot reasoners](https://md.research.elamur.ai/free/en/papers/kojima-2022-large-language-models-zero-shot/). In Advances in NeurIPS 2022.
13. Michal Kosinski. 2023. [Theory of mind may have spontaneously emerged in large language models](https://md.research.elamur.ai/free/en/papers/kosinski-2023-evaluating-large-language-models-theory/). arXiv preprint arXiv:2302.02083.
14. János Kramár, Tom Eccles, Ian Gemp, Andrea Tacchetti, Kevin R. McKee, Mateusz Malinowski, Thore Graepel, and Yoram Bachrach. 2022. Negotiation and honesty in artificial intelligence methods for the board game of Diplomacy. Nature Communications, 13(1):7214. [find](https://scholar.google.com/scholar?q=J%C3%A1nos+Kram%C3%A1r%2C+Tom+Eccles%2C+Ian+Gemp%2C+Andrea+Tacchetti%2C+Kevin+R%2E+McKee%2C+Mateusz+Malinowski%2C+Thore+Graepel%2C+and+Yoram+Bachrach%2E+2022%2E+Negotiation+and+honesty+in+artificial+intelligence+methods+for+the+bo)
15. David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. 2017. Mastering the game of go without human knowledge. Nature, 550(7676):354–359. [find](https://scholar.google.com/scholar?q=David+Silver%2C+Julian+Schrittwieser%2C+Karen+Simonyan%2C+Ioannis+Antonoglou%2C+Aja+Huang%2C+Arthur+Guez%2C+Thomas+Hubert%2C+Lucas+Baker%2C+Matthew+Lai%2C+Adrian+Bolton%2C+et+al%2E+2017%2E+Mastering+the+game+of+go+without+hu)
16. Fujio Toriumi, Hirotaka Osawa, Michimasa Inaba, Daisuke Katagami, Kosuke Shinoda, and Hitoshi Matsubara. 2017. AI wolf contest—development of game ai using collective intelligence—. In Computer Games, pages 101–115. Springer. [find](https://scholar.google.com/scholar?q=Fujio+Toriumi%2C+Hirotaka+Osawa%2C+Michimasa+Inaba%2C+Daisuke+Katagami%2C+Kosuke+Shinoda%2C+and+Hitoshi+Matsubara%2E+2017%2E+AI+wolf+contest%E2%80%94development+of+game+ai+using+collective+intelligence%E2%80%94%2E+In+Computer+Games%2C)
17. Ruibo Liu, Ruixin Yang, Chenyan Jia, Ge Zhang, Denny Zhou, Andrew M Dai, Diyi Yang, and Soroush Vosoughi. 2023. Training socially aligned language models in simulated human society. arXiv preprint arXiv:2305.16960. [find](https://scholar.google.com/scholar?q=Ruibo+Liu%2C+Ruixin+Yang%2C+Chenyan+Jia%2C+Ge+Zhang%2C+Denny+Zhou%2C+Andrew+M+Dai%2C+Diyi+Yang%2C+and+Soroush+Vosoughi%2E+2023%2E+Training+socially+aligned+language+models+in+simulated+human+society%2E+arXiv+preprint+arX)
18. Danqing Wang and Lei Li. 2023. Learn from mistakes through cooperative interaction with study assistant. arXiv preprint arXiv:2305.13829. — [arXiv:2305.13829](https://arxiv.org/abs/2305.13829) [find](https://scholar.google.com/scholar?q=Danqing+Wang+and+Lei+Li%2E+2023%2E+Learn+from+mistakes+through+cooperative+interaction+with+study+assistant%2E+arXiv+preprint+arXiv%3A2305%2E13829%2E+%E2%80%94+%5BarXiv%3A2305%2E13829%5D%28https%3A%2F%2Farxiv%2Eorg%2Fabs%2F2305%2E13829%29)
19. OpenAI. 2022. Introducing ChatGPT. (Accessed on Jun 18, 2023). [find](https://scholar.google.com/scholar?q=OpenAI%2E+2022%2E+Introducing+ChatGPT%2E+%28Accessed+on+Jun+18%2C+2023%29%2E)
20. OpenAI. 2023. [GPT-4 technical report](https://md.research.elamur.ai/free/en/papers/openai-2023-gpt-4-technical-report/). arXiv preprint arXiv:2303.08774.
21. Tianhe Wang and Tomoyuki Kaneko. 2018. Application of deep reinforcement learning in werewolf game agents. In TAAI 2018, pages 28–33. IEEE. [find](https://scholar.google.com/scholar?q=Tianhe+Wang+and+Tomoyuki+Kaneko%2E+2018%2E+Application+of+deep+reinforcement+learning+in+werewolf+game+agents%2E+In+TAAI+2018%2C+pages+28%E2%80%9333%2E+IEEE%2E)
22. Hirotaka Osawa, Fujio Toriumi, Daisuke Katagami, Kosuke Shinoda, and Michimasa Inaba. 2014. Designing protocol of werewolf game: Protocol for inference and persuasion. The 24th Fuzzy, Artificial Intelligence, Neural Networks and Computational Intelligence. [find](https://scholar.google.com/scholar?q=Hirotaka+Osawa%2C+Fujio+Toriumi%2C+Daisuke+Katagami%2C+Kosuke+Shinoda%2C+and+Michimasa+Inaba%2E+2014%2E+Designing+protocol+of+werewolf+game%3A+Protocol+for+inference+and+persuasion%2E+The+24th+Fuzzy%2C+Artificial+Intel)
23. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed H Chi, Quoc V Le, Denny Zhou, et al. 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/). In Advances in NeurIPS 2022.
24. Joon Sung Park, Joseph C O’Brien, Carrie J Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. arXiv preprint arXiv:2304.03442. [find](https://scholar.google.com/scholar?q=Joon+Sung+Park%2C+Joseph+C+O%E2%80%99Brien%2C+Carrie+J+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+arXiv+preprint+arXiv%3A230)
25. Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. 2023a. [Visual ChatGPT: Talking, drawing and editing with visual foundation models](https://md.research.elamur.ai/free/en/papers/wu-2023-visual-chatgpt-talking-drawing-editing/). arXiv preprint arXiv:2303.04671.
26. Nils Reimers and Iryna Gurevych. 2019. [Sentence-BERT: Sentence embeddings using Siamese BERT-networks](https://md.research.elamur.ai/free/en/papers/reimers-2019-sentence-bert-sentence-embeddings-siamese/). In Proceedings of EMNLP-IJCNLP 2019, pages 3982–3992, Hong Kong, China. Association for Computational Linguistics.
27. Yuxiang Wu, Zhengyao Jiang, Akbir Khan, Yao Fu, Laura Ruis, Edward Grefenstette, and Tim Rocktäschel. 2023b. ChatArena: Multi-agent language game environments for large language models. [https://github.com/chatarena/chatarena](https://github.com/chatarena/chatarena). [↗](https://github.com/chatarena/chatarena)
28. Hong Ri, Xiaohan Kang, Mohd Nor Akmal Khalid, and Hiroyuki Iida. 2022. The dynamics of minority versus majority behaviors: A case study of the mafia game. Information, 13(3). [find](https://scholar.google.com/scholar?q=Hong+Ri%2C+Xiaohan+Kang%2C+Mohd+Nor+Akmal+Khalid%2C+and+Hiroyuki+Iida%2E+2022%2E+The+dynamics+of+minority+versus+majority+behaviors%3A+A+case+study+of+the+mafia+game%2E+Information%2C+13%283%29%2E)
29. Shunyu Yao, Jeffrey Zhao, Dian Yu, 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 NeurIPS 2022 Foundation Models for Decision Making Workshop.
30. Natalie Shapira, Mosh Levy, Seyed Hossein Alavi, Xuhui Zhou, Yejin Choi, Yoav Goldberg, Maarten Sap, and Vered Shwartz. 2023. Clever hans or neural theory of mind? Stress testing social reasoning in large language models. arXiv preprint arXiv:2305.14763. [find](https://scholar.google.com/scholar?q=Natalie+Shapira%2C+Mosh+Levy%2C+Seyed+Hossein+Alavi%2C+Xuhui+Zhou%2C+Yejin+Choi%2C+Yoav+Goldberg%2C+Maarten+Sap%2C+and+Vered+Shwartz%2E+2023%2E+Clever+hans+or+neural+theory+of+mind%3F+Stress+testing+social+reasoning+in+l)
31. Denny Zhou, Nathanael Schärli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc V Le, and Ed H. Chi. 2023. [Least-to-most prompting enables complex reasoning in large language models](https://md.research.elamur.ai/free/en/papers/al-2022-least-most-prompting-enables-complex/). In ICLR 2023.
32. Hisaichi Shibata, Soichiro Miki, and Yuta Nakamura. 2023. Playing the werewolf game with artificial intelligence for language understanding. arXiv preprint arXiv:2302.10646. — [arXiv:2302.10646](https://arxiv.org/abs/2302.10646) [find](https://scholar.google.com/scholar?q=Hisaichi+Shibata%2C+Soichiro+Miki%2C+and+Yuta+Nakamura%2E+2023%2E+Playing+the+werewolf+game+with+artificial+intelligence+for+language+understanding%2E+arXiv+preprint+arXiv%3A2302%2E10646%2E+%E2%80%94+%5BarXiv%3A2302%2E10646%5D%28https)
33. Noah Shinn, Beck Labash, and Ashwin Gopinath. 2023. [Reflexion: an autonomous agent with dynamic memory and self-reflection](https://md.research.elamur.ai/free/en/papers/shinn-2023-reflexion-language-agents-verbal-reinforcement/). arXiv preprint arXiv:2303.11366.
34. David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur Guez, Marc Lanctot, Laurent Sifre, Dharshan Kumaran, Thore Graepel, et al. 2018. A general reinforcement learning algorithm that masters chess, shogi, and go through self-play. Science, 362(6419):1140–1144. [find](https://scholar.google.com/scholar?q=David+Silver%2C+Thomas+Hubert%2C+Julian+Schrittwieser%2C+Ioannis+Antonoglou%2C+Matthew+Lai%2C+Arthur+Guez%2C+Marc+Lanctot%2C+Laurent+Sifre%2C+Dharshan+Kumaran%2C+Thore+Graepel%2C+et+al%2E+2018%2E+A+general+reinforcement+lear)

*Updated 24 October 2026*