(Ulsan = Yonhap News) Reporter Kim Yong-tae = A technology has been developed that can train AI to not believe what other artificial intelligence (AI) says, and to infer and judge hidden intentions.
On the 1st, the Ulsan National Institute of Science and Technology (UNIST) announced that Professor Han Seung-yeol's team at the Graduate School of Artificial Intelligence developed 'MARBO (Multi-Agent Relational Belief Optimization)', which trains a large language model AI to directly infer the role and intention of the other party and speak and act according to that judgment.
According to the research team, multiple AI agents may intentionally deceive each other when they cooperate or compete.
Marbo is a learning method that teaches AI to identify the other person's hidden role and intentions in such situations, and to act and speak in accordance with its own goals.
This learning method first teaches you to infer what role the other person plays and who you are on the same side with based on accumulated conversations and actions.
This assumption about the other party's identity and relationship is called 'relational belief', and the combination of AIs participating together is also randomly changed to experience various situations.
In addition, a positive evaluation is given only when the opponent's identity is accurately judged and actions that are advantageous to the goal are selected, so that the child learns to make such judgments and actions even in similar situations.
In existing learning methods, even if you reasoned incorrectly, if you happen to get a favorable result, you are given a good evaluation and the choice is repeated, but Marbo can filter out this kind of thing.
Teach children to consider what others think of them when speaking. Don't just judge whether the words are appropriate for the situation, but also check whether the other person's judgment after hearing the words has changed in your favor.
The research team verified the performance of the learning method by having AI play against each other in 'Werewolf' and 'Among Us', a type of mafia game.
The mafia game is a social reasoning game in which each participant has to guess the hidden identity of the other player in a situation where each participant knows different information and may even lie. In the 'Werewolf' and 'Among Us' games, you must find a werewolf hidden among the citizens and a traitor who interferes with the spaceship mission, respectively.
When the research team trained a language model with 7 billion parameters using Marbo and then played the werewolf game, the average win rate against five types of opponent AI was 65.6%.
This is 24 percentage points higher than the 41.6% when trained using the existing learning method. It was higher than the 64.0% win rate recorded when training a model with twice the parameters using the existing learning method.
In the case of the Among Us game, the win rate for the model with about 2 billion parameters increased from 36.4% to 49.0%, and for the model with about 4 billion parameters, the win rate increased from 40.0% to 59.0%.
Professor Seung-Yeol Han said, “It is a technology that enables proper judgment by inferring the role and relationship of the other party even in situations where there is incomplete or deceptive information.” He added, “In the future, it can be expanded into a basic technology that increases system reliability and adaptability where multiple agents, such as multi-robots, autonomous drones, and human-AI collaboration, must understand each other’s status and intentions and cooperate.”
The research results were revealed at 'EMNLP' (Empirical Methods in Natural Language Processing), an international academic conference in the field of natural language processing. This year's competition was held in Budapest, Hungary from October 24th to 29th.
The research was supported by the Ministry of Science and ICT's Information and Communications Planning and Evaluation Institute and the National Research Foundation of Korea.
