Key Points
- Researchers at UNIST developed 'MARBO', an AI training method that enables large language models to infer hidden roles and motives of rival agents instead of trusting them blindly.
- Tested on social deduction games like Werewolf and Among Us, the method boosted a 7-billion parameter model's win rate to 65.6%, outperforming conventional models twice its size.
- The study taps into South Korea's vibrant culture of 'Mafia' parlor games to advance multi-agent AI cooperation and fraud detection for real-world robotics and drone systems.
Summary South Korean artificial intelligence researchers have taught AI agents how to play 'Mafia' games like Werewolf and Among Us—and more importantly, how to detect when other bots are lying to them.
A research team led by Professor Han Seung-ryeol at the Ulsan National Institute of Science and Technology (UNIST) Artificial Intelligence Graduate School revealed 'MARBO' (Multi-Agent Relational Belief Optimization). This training framework specifically instructs Large Language Models (LLMs) to infer the hidden roles, alliances, and ulterior motives of other agents rather than taking statements at face value, adapting their speech and actions accordingly.
Social deduction and parlor games—widely known in South Korea under the umbrella term 'Mafia games' (마피아 게임)—have long been a cornerstone of Korean pop culture, variety shows, and youth gatherings. In games like Werewolf or the hit multiplayer game Among Us, participants hold asymmetric information and must use deception, persuasion, and behavioral cues to unmask traitors while concealing their own true identities. By placing AI models in these competitive conversational settings, the UNIST team tested whether autonomous agents could navigate complex psychological maneuvering.
MARBO trains models by building 'relational beliefs' based on past dialogue and behavior, while constantly randomizing participating AI combinations to simulate diverse, unpredictable dynamics. Unlike conventional reinforcement learning methods that accidentally reward lucky outcomes even after faulty deductions, MARBO only rewards an agent when it accurately deduces another's identity and chooses strategically advantageous actions. Furthermore, agents are trained to consider how they are perceived by others, evaluating whether their own remarks actively persuade their counterparts.
The experimental results demonstrated a substantial leap in strategic gameplay. When a 7-billion parameter language model was trained with MARBO and pitted against five different AI opponents in Werewolf, its average win rate reached 65.6%—a 24 percentage-point surge from the 41.6% achieved via standard training. Remarkably, this outperformed a model twice its size trained under traditional methods (64.0%). Similar dramatic improvements were logged in Among Us benchmarks.
Presented at the international Empirical Methods in Natural Language Processing (EMNLP) conference in Budapest, the research marks an important milestone. Beyond gaming, researchers note that teaching AI to interpret deceit and hidden intentions is a foundational step for future multi-robot coordination, autonomous drones, and dependable human-AI collaboration in real-world scenarios where information is incomplete or intentionally misleading.
Sponsored · Ad