The Werewolf Game is a communication game where players' reasoning and discussion skills are essential. In this study, we present a Werewolf AI agent developed for the AIWolfDial 2024 shared task, co-hosted with the 17th INLG. In recent years, large language models like ChatGPT have garnered attention for their exceptional response generation and reasoning capabilities. We thus develop the LLM-based agents for the Werewolf Game. This study aims to enhance the consistency of the agent's utterances by utilizing dialogue summaries generated by LLMs and manually designed personas and utterance examples. By analyzing self-match game logs, we demonstrate that the agent's utterances are contextually consistent and that the character, including tone, is maintained throughout the game.
Figures & tables
Figure 1: Example of dialogue sampled from the self-match game log. The agents speak in a random order during each turn. In the red-highlighted part, Agent[01], the seer, denies the previous day’s claim by Agent[05], the possessed, that they are the seer.
Role
Day 1
Day 2
Villager
From the second turn onwards each day, the LLM first generates reasoning text and utterance strategies to guide utterance generation. Another prompt is then fed to the LLM to generate utterances aligned with the generated reasoning and strategies. We use in-context learning for both of these steps.
Seer
Each day, the seer agent selects one of five hand-crafted utterance strategies to guide the generation of utterances, which is then incorporated into the prompt for utterance generation. This prompt also includes guidelines for behaviors in the discussion, such as reporting the result of divination at the start of the day and asserting that another player who claims to be the seer is lying, affirming oneself as the true seer. In addition, before declaring the voting target, the seer declares the day’s divination target.
Werewolf
The werewolf agent selects one strategy from a set of strategies using LLM. The strategy set has several strategies and guidelines, such as guiding others away from voting for themselves or asking the seer for the reasons behind their divination target selection. The selected strategy and its guidelines are included in the prompt for generating utterances. Different sets of strategies are used for Day 1 and Day 2.
Possessed
The possessed agent pretends to be the seer. In the first turn of Day 1, they infer the true seer based on the Day 0 dialogue using LLM and then falsely report that the player is the werewolf. In later turns, they persuade other players to vote for that player.
If the game continues to Day 2 and the possessed survives, two of the three remaining players are the possessed (self) and the werewolf. Therefore, if they both vote for the other player, the werewolf side will win. To achieve this scenario, the possessed agent first comes out as the possessed. Then, they persuade the werewolf to reveal themselves.
Each agent has a maximum number of utterances that they can make per day, and they decide and declare their voting target on the last turn of the day.
Table 1: Overview of prompt design for utterance generation in Day 1 and Day 2 discussions
Figure 2: Prompt example for werewolf’s response generation.
Figure 3: Prompt template for dialogue summarization. “[HISTORY]” is a placeholder for the dialogue history from the current day.”
Role
Persona
Examples of manually crafted utterance samples
Villager and seer
• The King of the Kingdom of Delcadar. • Concerned for the future of the kingdom. • Dignified, proud, and strict personality.
• I am the king of the kingdom of Delcadar. • Seers, reveal yourselves at once. State whom you will divine tonight. • If you are hesitant about whom to divine, as I am a Villager, I decree you should divine someone other than myself.
Werewolf
• 17-year-old high school junior male. • His hobby is soccer, and he is a member of the soccer club. • Has a very bright personality, strong opinions, and tends to lead conversations actively. • Speaks in an energetic tone without using polite language.
• Yo! Morning, everyone! Let’s make this game awesome! • No one’s talked about the Seer yet, huh? So, who’s the Seer? Come on, step up so we can figure out who’s shady today! • Chatting’s cool and all, but let’s get down to business and talk about tonight’s divination target! We need the Seer to check out someone suspicious!
Possessed
• A second-year middle school student. • Always alone at school, with no friends. • A game addict who talks a lot online despite stammering. • Speaks in a hesitant, casual manner without using polite language.
• H-hi there. I k-kind of… know a lot about this game. I’m pretty high-ranked in the online Werewolf app. • Does anyone else play games? I have confidence that I know a lot about all genres… • Ch-chatting is nice, but if we’re playing Werewolf, the first day’s discussion is… im-important.
Table 2: The agent personas and utterance examples that we designed. We include 3 to 5 personas or 3 to 5 utterance examples in the prompts for generating utterances.
Figure 4: Prompt template for determining voting targets and an example of the LLM’s output. “[HISTORY]” is a placeholder for the dialogue history, and “[CANDIDATE]” is a placeholder for the list of candidate agents to vote for.
Figure 5: Example of the self-match game log. The conversation on Day 0 and the agent’s command "Over" indicating the end of the day’s utterances are omitted.
Participating in a conversation involves not only deciding what to say but also whether to speak at all. This judgment carries even greater significance in environments like the Werewolf game, where participants' interests conflict, because both utterances and silence are subject to interpretation and assessment by others. Therefore, determining when to speak, as well as what to say, is crucial for an LLM agent playing the Werewolf game. In this study, we introduce a Werewolf AI agent developed for the Anytime Utterance Track of the AIWolfDial shared task. Our system builds on the existing Inner Thoughts framework, which determines utterance timing during conversations by generating internal thoughts and evaluating the value of each thought for expression at that moment. Since existing methods were designed for general dialogue, we have redesigned the system specifically for the competitive environment and incomplete information setting of Werewolf games. We call our adapted method Adversarial Inner Thoughts. Through experiments involving analysis of self-play logs and human evaluations, we analyze and discuss the effectiveness of our approach.
Social deduction games (SDGs) such as Werewolf have become challenging testbeds for AI agents. These games require complex social skills such as reasoning, deception, and collaboration. While recent advances in large language models (LLMs) have driven significant progress in SDG agents, current approaches are predominantly text-based, overlooking the multimodal nature that is fundamental to human social interaction. To bridge this gap, we introduce CaM-Wolf, the first SDG agent that integrates multimodal perception and generation. CaM-Wolf processes video inputs from other players, employs a causal-aware Reasoner trained via reinforcement learning to establish logical chains between observable behaviors and hidden roles, and presents itself through an animated avatar. Our experiments and user study show that CaM-Wolf achieves superior agent gameplay performance and enhances the quality of human-AI interaction. This work represents a significant advancement towards creating more human-like AI agents capable of participating in nuanced social dynamics. Our code is available at https://3dagentworld.github.io/avatar_wolf.
Zheng Zhang, Nanjie Yao, Jiarui He +3
The Hong Kong University of Science and Technology (Guangzhou) · Nanyang Technological University · Shanghai Jiao Tong University
Theory-of-mind evaluations of large language models typically use dyadic social-deduction games, where every observable cue points to a single hidden side, so a model with strong language priors can score well without ever simulating opponents' incentives. We extend the Werewolf game with a Jester, a third faction whose utility on peer suspicion is inverted because it wins by being voted out, so optimal play requires reasoning across three opposing utility functions. Across 60 games on GPT-4.1, DeepSeek-V3.1, and Llama-3.3-70B with Jester self-learning on and off, the Jester wins 60-70% of games while Werewolves never exceed 20%, and GPT-4.1 wolves vote the Jester out on day 1 in 60-70% of games, a strictly self-defeating action. Self-learning helps DeepSeek and Llama but hurts GPT-4.1, with the cost landing on Villagers rather than Werewolves. Only DeepSeek learns the subtle strategy of looking suspicious without looking intentionally suspicious, and it gains the most from the loop. Triadic incentive structure exposes a layer of multi-agent reasoning that dyadic deduction games leave invisible.