CaM-Wolf: Multimodal AI Agent Excels in Social Deduction Games.

Zheng Zhang, Nanjie Yao, Jiarui He, Deheng Ye, Peilin Zhao, Hao Wang· July 31, 2026 View original

Key takeaways

  • CaM-Wolf is the first multimodal AI agent for social deduction games.
  • It uses video input and causal reasoning to infer hidden roles.
  • The agent enhances human-AI interaction and achieves superior gameplay.
  • This work advances AI's ability to handle complex social dynamics.

Who benefits

GamingCustomer ServiceVirtual AssistantsEdTechEntertainment

Summary

CaM-Wolf is the first social deduction game agent to integrate multimodal perception and generation, processing video inputs and using a causal-aware Reasoner to infer hidden roles. It enhances human-AI interaction and achieves superior gameplay performance in games like Werewolf.

Social deduction games, such as Werewolf, present significant challenges for AI due to their reliance on complex social skills like reasoning, deception, and collaboration. While large language models (LLMs) have improved AI agents in these games, existing approaches are largely text-based, neglecting the crucial role of multimodal interaction in human social dynamics. To bridge this gap, researchers introduce CaM-Wolf, an innovative social deduction game agent that incorporates both multimodal perception and generation. CaM-Wolf processes video feeds from other players, utilizing a reinforcement learning-trained, causal-aware Reasoner to establish logical connections between observable behaviors and hidden roles. It then presents itself through an animated avatar, making interactions more human-like. Experiments and user studies confirm that CaM-Wolf not only achieves superior gameplay performance but also significantly improves the quality of human-AI interaction. This development marks a substantial step towards creating AI agents that can engage in more nuanced and human-like social dynamics.

Why it matters

This advancement demonstrates AI's growing capability in understanding and participating in complex social interactions, which has implications for virtual assistants, customer service, and even training simulations.

How to implement this in your domain

  1. 1Explore the principles of multimodal AI and causal reasoning for enhancing interactive AI applications.
  2. 2Investigate how multimodal input (e.g., video, audio, text) can improve the performance of existing conversational AI systems.
  3. 3Consider developing animated avatars or visual representations for AI agents to improve user engagement and perceived social intelligence.
  4. 4Pilot AI agents in simulated social environments to test their ability to reason, deceive, or collaborate effectively.

Original post by Zheng Zhang, Nanjie Yao, Jiarui He, Deheng Ye, Peilin Zhao, Hao Wang

"arXiv:2607.26393v1 Announce Type: new Abstract: 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…"

View on X

Originally posted by Zheng Zhang, Nanjie Yao, Jiarui He, Deheng Ye, Peilin Zhao, Hao Wang on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Research

AI ResearchAI Engineering & DevTools

New Framework Improves Partial Multi-View Clustering Performance.

DAS-PMVC is a novel framework for partial multi-view clustering that addresses view asymmetry and irrelevant samples by leveraging dual alignment and structure enhancement. It uses anchor graph structure alignment, structure-enhanced feature learning, and a dual alignment strategy to achieve superior clustering performance on various datasets.

Shubin Ma, Liang Zhao, Chuanye He, Zhenjiao Liu, Liang Zou, Lin Yuanbo Wu, Yu ShaoJul 31, 2026
AI Engineering & DevToolsAI Research

Dual Teachers Improve Adversarial Robustness and Accuracy.

This work extends Information Bottleneck Distillation (IBD) by introducing a "clean teacher" alongside a robust teacher to improve the robustness/accuracy tradeoff against adversarial attacks. The proposed method transfers features from both teachers to a student model, achieving better clean accuracy while maintaining adversarial robustness, outperforming original IBD and competing with state-of-the-art approaches.

Vincent Ryusuke Takahashi, Yoshinari Takeishi, Jun'ichi Takeuchi, Kave SalamatianJul 31, 2026
AI Engineering & DevToolsAI Research

Dynamic Batch Sizes Improve Large Language Model Training Efficiency.

This paper proposes a new approach to deep learning dynamics, deriving joint scaling laws for loss based on both learning rate and batch size schedules. It introduces an optimal dynamic batch size schedule that consistently outperforms static batch size baselines, highlighting its importance for large language model training.

Jiaxiang Li, Zhiqi Bu, Shiyun XuJul 31, 2026