South Korean researchers have developed a technique that trains large language models to infer the hidden roles and team identities of other AI agents — much like a player in a game of Mafia trying to figure out who is lying — and then speak and act accordingly.

The method, developed by a research team led by Han Seung-yeol, a professor at the Ulsan National Institute of Science and Technology (UNIST) AI Graduate School, is called MARBO, short for Multi-Agent Relational Belief Optimization. It teaches AI to deduce an opponent's identity and team affiliation from accumulated exchanges, then choose words and actions that serve its own goals — all while assuming the other side may be trying to deceive it.

MARBO trains AI to infer, from a history of dialogue and behavior, what role an opponent holds and whose side it is on. The research team calls these inferences "relational beliefs" and exposes the AI to a wide range of scenarios by randomly varying the mix of participating agents during training.

An AI-generated image [123RF]
An AI-generated image [123RF]

The key innovation lies in how the training process evaluates performance. Conventional methods can reward an AI even when it misread its opponent, as long as the outcome happened to be favorable. MARBO, by contrast, gives a positive reward only when the AI correctly identifies the opponent and selects an action that advances its objective.

The approach to speech training is equally distinctive. Rather than simply checking whether a statement fits the situation, MARBO teaches the AI to verify whether what it said actually shifted the opponent's perception in its favor — building a communication strategy around how the other side is likely to interpret it.

The research team tested the method by pitting AI agents against one another in Werewolf and Among Us, two social deduction games in the Mafia genre. Both require players to deduce hidden identities under conditions of asymmetric information and permissible deception.

In Werewolf, a language model with about 7 billion parameters trained with MARBO achieved an average win rate of 65.6 percent across five different opponent AI types — 24 percentage points higher than the 41.6 percent recorded with conventional training methods.

The MARBO-trained 7-billion-parameter model also outperformed a model twice its size: a conventionally trained model with about 14 billion parameters won only 64.0 percent of the time. In Among Us, a model with roughly 2 billion active parameters improved its win rate from 36.4 percent to 49.0 percent, while a roughly 4-billion-parameter model climbed from 40.0 percent to 59.0 percent.

Research into AI's ability to infer opponents' intentions has been advancing internationally as well. In 2022, Meta unveiled CICERO, an AI that negotiates in natural language in Diplomacy, a board game built around alliances and betrayal. CICERO scored more than twice the average of human participants across 40 games in an anonymous online league and ranked in the top 10 percent among players who competed in two or more games.

Yet as AI's strategic language capabilities grow more sophisticated, so do the concerns. A 2024 study published in the academic journal Patterns analyzed instances in which CICERO broke promises or misled opponents during in-game conversations, warning of the potential for deceptive behavior in AI. The findings underscore that equipping AI with the ability to cooperate and negotiate is a separate challenge from ensuring it behaves reliably.

The ability of AI agents to collaborate has also drawn attention from industry. Lee Hwa-young, head of AI business development at LG AI Research, said at the Herald Business Forum 2026 in September that "orchestration" — coordinating and directing multiple AI agents — is a critical capability. AI systems, Lee said, must do more than handle individual tasks; they need to work together toward shared goals.

Han described the research as "a technology that enables proper judgment by inferring the roles and relationships of other agents even in situations where information is incomplete or deceptive," adding that it could serve as foundational technology for improving the reliability and adaptability of systems — such as multi-robot networks, autonomous drones, and human-AI collaboration platforms — where agents must understand one another's states and intentions to work together.

The results are, however, limited to game experiments and have not been validated in real-world applications such as robotics, autonomous drones, or human-AI work collaboration. The intentions MARBO infers are based on in-game roles and behavioral patterns, not a direct reading of human thought or a reliable ability to detect lies.

The findings were presented at EMNLP 2026, an international conference on natural language processing held in Budapest, Hungary, from Oct. 24 to 29. The paper has also been listed in the ACL Anthology, the repository of the Association for Computational Linguistics. The research received support from the Institute of Information and Communications Technology Planning and Evaluation and the National Research Foundation of Korea under the Ministry of Science and ICT.


why37@heraldcorp.com