Enhancing Consistency of Werewolf AI through Dialogue Summarization and Persona Information

This paper presents an LLM-based Werewolf AI agent for the AIWolfDial 2024 shared task that improves utterance consistency and character maintenance by leveraging dialogue summaries and manually designed personas.

The Gist

Imagine a high-stakes game of Mafia (or Werewolf) played over a text chat. In this game, some players are "villagers" trying to find the bad guys, while others are "werewolves" and a "possessed" person trying to trick everyone. The catch? You can't see who is who. You have to figure it out just by listening to what people say, watching how they act, and remembering everything that happened yesterday.

The researchers in this paper built a team of AI players to compete in this game. Their goal was to make these AI players act like real humans: smart, consistent, and with distinct personalities.

Here is how they did it, explained with some everyday analogies:

1. The Problem: The "Too Much Information" Overload

Imagine you are playing a game that lasts for three days. By Day 2, you have to remember every single thing everyone said on Day 1, plus the new stuff on Day 2. If you tried to read the entire transcript of a 3-day conversation every time you wanted to speak, your brain would get tired, you'd get confused, and you might forget the most important clues.

The AI's Solution: The "Cheat Sheet" (Dialogue Summarization)
Instead of forcing the AI to read the entire 50-page transcript of the game history every time it speaks, the researchers gave it a condensed cheat sheet.

• How it works: At the end of every day, the AI uses its own brain (a Large Language Model) to write a short summary of what happened. It highlights who claimed to be the "Seer" (the detective), who was suspicious, and who got voted out.
• The Benefit: When the game starts on Day 2, the AI doesn't need to re-read the whole book. It just reads the "Cliff's Notes" summary. This keeps the AI fast, saves money on computer power, and helps it remember the important clues without getting lost in the noise.

2. The Problem: The "Chameleon" Effect

Have you ever talked to a friend who suddenly changes their personality based on who they are talking to? One minute they are a shy librarian, and the next they are a loud party animal because they are trying to fit in. In a game like Werewolf, if your AI agent changes its tone or personality every time someone else speaks, the other players will know it's a robot. It breaks the immersion.

The AI's Solution: The "Character Sheet" (Persona Design)
To stop the AI from acting like a chameleon, the researchers gave each AI a strict character profile, like a role-playing game character sheet.

• The Villager/Seer: They are the "King of a Kingdom." They speak with dignity, pride, and strictness. They don't use slang.
• The Werewolf: They are a "17-year-old soccer player." They are energetic, loud, use slang, and hate being polite.
• The Possessed: They are a "shy, stuttering gamer." They talk nervously and act awkward.
• The Magic: The researchers didn't just tell the AI what to say; they gave it examples of how these characters talk. Every time the AI speaks, it checks its character sheet to make sure it sounds like the "Soccer Player" or the "King," not like a generic robot. This makes the AI feel like a real person with a consistent soul.

3. The Problem: The "Oops, I Contradicted Myself" Moment

In a complex game, it's easy to make a mistake. You might say, "I think Player A is the werewolf," and then five minutes later vote for Player B without explaining why. This looks suspicious and untrustworthy.

The AI's Solution: The "Thinking Aloud" Strategy (Chain-of-Thought)
Before the AI makes a big move (like voting someone out or guessing a role), the researchers made it think out loud first.

• How it works: The AI is forced to write down its reasoning step-by-step before it types its final answer. It asks itself: "Why do I suspect Player A? What did they say yesterday? Does this match my plan?"
• The Benefit: This acts like a safety net. It ensures the AI's final vote matches its earlier arguments, making it look much more logical and human.

The Result: A Winning Team

The researchers tested their AI by having them play against each other. The results showed that:

1. They remembered everything: Even on Day 2, the AI could recall specific details from Day 1 because of the "Cheat Sheet" summaries.
2. They stayed in character: The "King" sounded royal the whole time, and the "Soccer Player" stayed energetic, even when the game got stressful.
3. They were consistent: They didn't flip-flop on their votes or contradict their own stories.

In a nutshell: The researchers taught the AI to be a better game player by giving it a summary of the past (so it doesn't forget), a strict personality guide (so it doesn't act weird), and a thinking process (so it doesn't make mistakes). This makes the AI feel less like a computer program and more like a clever, consistent human player.

Technical Summary

Here is a detailed technical summary of the paper "Enhancing Consistency of Werewolf AI through Dialogue Summarization and Persona Information."

1. Problem Statement

The paper addresses the challenges of developing AI agents for the Werewolf Game (specifically for the AIWolfDial 2024 shared task). The game is a complex, incomplete-information communication game requiring high-level reasoning, strategic deception, and natural language interaction among five players (Seer, Werewolf, Possessed, and two Villagers).

The authors identify three primary technical hurdles in applying Large Language Models (LLMs) to this domain:

• Contextual Consistency: Agents must maintain consistent claims and logic across multiple days of dialogue. Long dialogue histories often lead to hallucinations or contradictions when fed entirely into LLM prompts.
• Character Coherence: Agents must maintain a distinct persona (tone, personality, and speech patterns) throughout the game, resisting the influence of other players' tones or generating contradictory character traits.
• Efficiency Constraints: Feeding the entire game history (including greetings and redundant utterances) into LLM prompts increases API costs, generation latency, and exceeds input token limits, while introducing noise that hinders decision-making.

2. Methodology

The proposed system utilizes a modular LLM-based architecture designed to handle the specific constraints of the Werewolf Game. The methodology focuses on three core techniques:

A. Dialogue Summarization for Context Management

To address the limitations of long context windows and information redundancy:

• Mechanism: Instead of feeding the raw, full dialogue history from previous days into the prompt, the system employs an LLM to generate a structured summary of the previous day's interactions at the end of each day.
• Content: The summary captures critical game state information, including:
  • Claims made by players (e.g., who claimed to be the Seer).
  • Divination results.
  • Suspicions and reasoning provided by players.
  • Contradictory statements.
• Application: This condensed summary is injected into the prompt for the current day's turn, allowing the agent to retain crucial historical context without the token overhead of the full transcript.

B. Persona-Driven Utterance Generation

To ensure character consistency and diverse expression:

• Hand-Crafted Personas: The authors manually designed specific profiles for each role (Villager, Seer, Werewolf, Possessed). These include demographic details, personality traits, and specific speech styles (e.g., a "King" persona for the Seer vs. a "high school soccer player" for the Werewolf).
• In-Context Learning: The prompts include these persona descriptions alongside few-shot examples of manually crafted utterances that exemplify the desired tone and vocabulary.
• Goal: This forces the LLM to adhere to a specific character voice, preventing the "generic AI" tone and ensuring the agent does not inadvertently adopt the speech patterns of opponents.

C. Chain-of-Thought (CoT) Reasoning for Action Decisions

To improve the logical consistency of game actions (voting, divination, attacks):

• Mechanism: The system uses Chain-of-Thought prompting (specifically zero-shot CoT) for decision-making processes.
• Process: Before generating a final action (e.g., "Vote for Agent X"), the LLM is instructed to output a step-by-step reasoning trace. This trace analyzes the dialogue history, evaluates suspicions, and aligns the decision with the agent's previous claims.
• Role-Specific Strategies:
  • Seer: Uses CoT to select divination targets based on conflicting claims.
  • Werewolf: Uses CoT to select attack targets and voting strategies that minimize suspicion.
  • Possessed: Uses CoT to infer the true Seer and formulate a strategy to mislead others.

D. Prompt Architecture

The prompt structure for utterance generation (e.g., for the Werewolf role) integrates:

1. Task Description & Rules.
2. Agent Persona & Utterance Examples.
3. Selected Speaking Strategy (chosen from a set of hand-crafted strategies).
4. Dialogue Summary (from previous days).
5. Current Day's Dialogue History.

3. Key Contributions

1. Novel Agent Framework: Development of four distinct AI agents (Villager, Seer, Werewolf, Possessed) specifically tuned for the Werewolf Game using LLMs.
2. Hybrid Consistency Approach: The integration of LLM-generated dialogue summaries (for memory efficiency and context retention) with manually crafted personas (for character stability). This dual approach addresses both logical consistency and stylistic coherence.
3. Strategic Reasoning: Implementation of Chain-of-Thought prompting to ensure that game actions (votes, attacks) are logically derived from the dialogue context and consistent with the agent's prior statements.
4. Empirical Validation: A case study demonstrating a full 5-player self-match game, proving the system's ability to maintain consistency over multiple days.

4. Results

The authors evaluated their system through a self-match game log (where all agents were controlled by their system). Key findings include:

• Contextual Consistency: Agents successfully retained information from Day 0 and Day 1 when playing on Day 2. For example, an agent correctly identified a Seer's claim from a previous day despite the intervening dialogue, proving the effectiveness of the summarization technique.
• Character Maintenance: Agents maintained their assigned personas throughout the game.
  • Example: The "King" persona (Seer) consistently used dignified, formal language.
  • Example: The "High Schooler" persona (Werewolf) maintained an energetic, informal tone without polite markers.
• Action-Claim Alignment: Agents demonstrated high consistency between their verbal claims and their final actions. If an agent claimed a player was suspicious during the discussion, they consistently voted for that player at the end of the day.
• Role Differentiation: The system successfully handled the complex dynamics of the "Possessed" role, which requires pretending to be a Seer while secretly aiding the Werewolf, adapting strategies dynamically as the game progressed.

5. Significance

This paper makes a significant contribution to the field of Multi-Agent Dialogue Systems and Game AI:

• Scalability: By utilizing dialogue summarization, the method offers a scalable solution for long-horizon dialogue tasks where full context retention is otherwise computationally prohibitive.
• Character Robustness: It demonstrates that manually curated personas combined with few-shot learning can effectively anchor LLM behavior, a crucial requirement for immersive role-playing and competitive social deduction games.
• Reasoning in Social Games: The work highlights that LLMs, when guided by Chain-of-Thought and structured prompts, can perform complex strategic reasoning in environments with hidden information and deception, moving beyond simple text generation to strategic gameplay.
• Benchmarking: The agents developed for AIWolfDial 2024 serve as a strong baseline for future research in LLM-based social deduction agents.