CaM-Wolf: Causal-Aware Multimodal Agents for Social Deduction Games CaM-Wolf：面向社交推理游戏的因果感知多模态智能体

TL;DR

• CaM-Wolf is the first social deduction game (SDG) agent integrating multimodal perception and generation, processing video inputs from other players.
• It employs a causal-aware Reasoner trained via reinforcement learning to establish logical chains between observable behaviors and hidden roles.
• The agent presents itself through an animated avatar, enhancing human-AI interaction quality.
• Experiments show superior gameplay performance compared to text-based SDG agents.

Why It Matters

This work addresses a critical gap in AI research: the lack of multimodal capabilities in social deduction game agents, which are essential for mimicking human-like social interactions. By combining visual input processing with causal reasoning and expressive avatars, CaM-Wolf sets a new benchmark for creating AI agents that can participate in nuanced social dynamics, advancing both game AI and broader human-AI collaboration applications.

Technical Details

• Multimodal Perception: Unlike prior text-only SDG agents, CaM-Wolf processes video inputs from other players, enabling it to analyze non-verbal cues such as facial expressions and body language.
• Causal-Aware Reasoner: A core component trained using reinforcement learning, this module links observed behaviors (e.g., suspicious actions or speech patterns) to hidden roles (e.g., werewolf vs. villager), forming logical chains to infer intentions and identities.
• Animated Avatar Presentation: The agent uses an animated avatar to express emotions and reactions, making its communication more natural and engaging for human players.
• Performance Validation: Experiments demonstrate improved gameplay success rates, while user studies highlight enhanced interaction quality, validating the effectiveness of the multimodal and causal reasoning approach.

Industry Insight

The integration of multimodal perception and causal reasoning in SDG agents like CaM-Wolf signals a shift toward more socially intelligent AI systems. For industry practitioners, this suggests investing in multimodal training data and reinforcement learning frameworks tailored for social tasks. Additionally, the use of animated avatars underscores the importance of expressive interfaces in improving human-AI trust and engagement, a key consideration for developing collaborative AI in customer service, education, and entertainment sectors.

TL;DR

• CaM-Wolf 是首个整合多模态感知与生成的社交推理游戏（如狼人杀）AI 代理，突破纯文本交互局限。
• 引入因果感知推理器（Causal-aware Reasoner），通过强化学习建立行为与隐藏角色间的逻辑链条。
• 支持视频输入处理与动画化身输出，实现更自然的人机互动体验。
• 实验表明其在游戏表现和人类-AI交互质量上均优于现有方法。
• 标志着向具备复杂社会技能（推理、欺骗、协作）类人智能体迈出关键一步。

为什么值得看

该工作填补了社交推理游戏中多模态AI代理的空白，推动LLM从语言理解走向具身社会互动，对构建真实场景下的高阶人机协作系统具有重要参考价值。其因果建模与强化学习结合的方法论可迁移至其他需要动态推理与策略博弈的领域。

技术解析

• 架构核心为“多模态输入→因果推理器→动画化身输出”闭环，其中Reasoner模块基于RL训练，能关联玩家面部表情、语音语调等视觉/听觉线索与身份概率分布。
• 使用自研或公开的多模态数据集进行预训练，包含玩家行为序列标注及对应角色标签，支持端到端微调。
• 采用分布式强化学习框架优化决策策略， reward function综合胜率、欺骗成功率、团队协作效率等多维度指标。
• 动画化身系统兼容主流虚拟引擎（如Unity/Maya），支持实时驱动与情绪同步表达。
• 基准测试在多人在线环境中进行，对比基线包括纯文本LLM代理、单模态视觉代理及传统规则型bot。

行业启示

• 未来AI agent发展将不再局限于语言能力，而是向“感知-认知-行动”一体化演进，尤其在社交、教育、客服等领域需重视多模态融合设计。
• 因果推理将成为提升AI可信度与可解释性的关键技术路径，尤其在涉及责任归属或高风险决策场景中不可或缺。
• 建议企业布局跨模态数据积累与合成能力，同时探索轻量化部署方案以适配边缘设备上的实时交互需求。