On November 1, the Ulsan National Institute of Science and Technology (UNIST) announced that a research team led by Professor Han Seung Yeol of the Graduate School of Artificial Intelligence has developed a training method called MARBO (Relational Trust Optimization between Agents).

This method helps large language models (LLMs) not only receive information but also infer the roles, stances and intentions of interacting objects, thereby choosing a way to communicate and act in accordance with the goal.

What's new about MARBO is the ability to help AI identify intentions that are not directly expressed through words or behavior. In an environment where many AI agents cooperate or compete, subjects can hide information, make false judgments, or intentionally deceive each other.

To do this, AI is trained to analyze conversation history and behavior to form assumptions about the identities, roles and relationships between objects. The groups of participating agents are randomly changed so that the AI ​​adapts to a variety of situations.

Notably, the system not only evaluates the final result but also considers the accuracy in the inference process. AI is only evaluated positively when it correctly identifies the object and chooses actions that are truly beneficial to the goal, instead of achieving favorable results randomly.

The research team tested the method by pitting multiple AI models against two social reasoning games, Werewolf and Among Us. These are games that require participants to detect secret roles, deduce intentions, and respond to deception.

As a result, the 7 billion parameter language model trained with MARBO achieved an average win rate of 65.6% in the game Werewolf when facing 5 types of AI, higher than the 41.6% level of the current training method. This result also exceeds the 64% level of the model with twice as many parameters but using the old method.

In the Among Us game, the winning rate of the 2 billion parameter model increased from 36.4% to 49%, while the 4 billion parameter model increased from 40% to 59%.

Professor Han Seung Yeol said MARBO helps AI infer roles and relationships between objects more accurately even when information is lacking or misleading. This technology has potential applications in multi-agent robotic systems, autonomous drones, and human-AI collaboration models.

The research results were announced at the International Conference on Experimental Methods in Natural Language Processing (EMNLP), taking place in Budapest (Hungary) from October 24 to 29, 2026. The research received support from the Institute for Information and Communication Planning and Evaluation and the National Research Foundation of Korea.

Maximum:

1500 characters There are currently no comments. Be the first to comment