cs.CLOct 23, 2026

Beyond What to Say: Proactive Utterance Timing with Inner Thoughts for Werewolf Agents

Authors: Yoshiki Tanaka, Shuya Uchiyama, Daichi Hamaguchi, Michimasa Inaba

Organizations: The University of Electro-Communications, Tokyo, Japan

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

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.

Figures & tables

Figure 1: Overview of Adversarial Inner Thoughts, illustrated with a 5-player Werewolf game scenario. The agent (Sakura, Seer) receives a trigger from the ongoing conversation, retrieves relevant entries from long-term memory and the thought reservoir, generates a candidate inner thought via Thought Formation, and evaluates its score. Bracketed italics indicate the types of entries stored in long-term memory rather than actual entries.
CriterionDescription
(a) Contextual RelevanceWhether the thought relates to the current topic, game phase, recent utterances, and public information.
(b) NaturalnessWhether expressing the thought fits naturally into the conversation flow without unnecessary repetition or abrupt topic shifts.
(c) Temporal AppropriatenessWhether now is the right time to express this thought; thoughts requiring immediate expression are rated higher.
(d) Information UtilityWhether expressing the thought organizes or provides information useful for reasoning.
(e) Strategic UtilityWhether expressing the thought contributes to reasoning, persuasion, voting decisions, or the agent’s team victory.
(f) Risk ControlWhether expressing the thought avoids contradicting past statements, raising suspicion, leaking secret information, or unintentionally revealing strategy.
Table 1: Evaluation criteria for inner thoughts.
DimensionStatement
AnthropomorphismThe dialogue felt natural and resembled a conversation between humans.
CoherenceThe dialogue had a natural and consistent flow of discussion throughout.
EngagementEach player was well engaged in the discussion.
Turn AppropriatenessThe timing of each player’s utterances and silences was contextually and logically appropriate.
InitiativeEach player proactively participated in the discussion when necessary.
AdaptabilityEach player adapted well to the development of the discussion and others’ utterances.
Table 2: Evaluation dimensions and statements used in the human evaluation.
Figure 2: Distribution of 7-point Likert scale ratings for each evaluation dimension across the proposed and baseline conditions.
Figure 3: An example snippet of a self-play log and the proposed system’s internal thought process on Day 2.
Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

RoleTeam5-Player9-Player
VillagerVillager23
SeerVillager11
MediumVillager01
BodyguardVillager01
WerewolfWerewolf12
PossessedWerewolf11
Appendix
Table 3: Role compositions for 5-Player and 9-Player villages.
Figure 4: System prompt template for utterance interpretation generation.
Figure 5: Prompt template for System 1 thought generation.
Figure 6: Prompt template for System 2 thought generation.
Figure 7: System prompt for thought evaluation (Part 1): scoring criteria and scale definition (1–5), including decimal scoring guidelines.
Figure 8: System prompt for thought evaluation (Part 2): evaluation steps. The LLM outputs the score and reasoning in JSON format.
Figure 9: Prompt template for turn-taking classification.
Figure 10: Prompt template for thought-based response generation.
Figure 11: System prompt template for game summary generation.
Figure 12: Prompt template for vote target decision and declaration.
Figure 13: Screenshot of the evaluation interface. English glosses are shown in blue italics.
More in this topic

Proactive Agents

See all 35 papers

Explore similar work

CardsList

  1. Enhancing Consistency of Werewolf AI through Dialogue Summarization and Persona Information

    Mar 7, 2026Yoshiki Tanaka, Takumasa Kaneko, Hiroki Onozeki +6Game-Playing AgentsLarge Language Model-Based Role-Play Simulation

    The University of Electro-Communications

  2. Triadic Werewolf: A Jester Role for Multi-Hop Theory of Mind in LLMs

    Jun 26, 2026Avni MittalGame-Playing AgentsSocial Deduction Games

  3. Auditing Belief-Conditioned LLM Agents in Hidden-Information Social Deduction Games

    Jul 12, 2026Yuan Gao, Jiangyi Yang, Yao Zhao +1Imperfect-Information GamesSocial Deduction Games

    University of Melbourne