AI That Detects Lies Unveiled... UNIST Develops 'MARBO' Psychological Reasoning Technology

An AI training method that helps AI survive among liars by reading its opponents

AI sees through the intentions behind "You're the Mafia, aren't you?"

Korean researchers have developed an artificial intelligence (AI) training technique that identifies the true intentions behind an opponent's lies and actions to find the best move.

A new AI learning method that filters out false information from conversation partners and reads the state of play and their intentions has demonstrated its performance in the Mafia game.

Korean researchers have developed an AI training technique that sees through opponents' lies and deceptive behavior and responds appropriately to the situation. The technique has demonstrated the ability to infer opponents' true intentions and social connections when information is incomplete.

A research team led by Professor Han Seungyeol at the Graduate School of Artificial Intelligence at UNIST (Ulsan National Institute of Science and Technology) announced on November 1 that it had developed MARBO, a large language model (LLM) training method that helps AI infer opponents' roles and objectives on its own.

Researchers (from left): Professor Han Seungyeol, researcher Hwang Yechan, and researcher Bae Sangjun. Photo provided by UNIST.

MARBO considers not only what opponents say but also their past actions. It determines who is allied with whom and what roles they have before deciding how to act.

The research team applied the technique to Werewolf and Among Us, social deduction games in which players try to identify participants hiding their identities. When MARBO was applied to an AI model with 7 billion parameters, the model achieved an average win rate of 65.6% in Werewolf.

That was 24 percentage points higher than the rate for existing models (41.6%). It also surpassed the 64.0% achieved by a model with twice as many parameters.

Instead of the existing approach, which rewarded accidental wins, the researchers improved accuracy by assigning weight only when sound judgment was combined with appropriate action.

First author and researcher Hwang Yechan said, "This is a new approach that incorporates learning about opponents' roles and relationships, rather than a technique limited to a specific game."

Professor Han Seungyeol said, "This technology enables sound judgment by inferring opponents' roles and relationships even when information is incomplete or deceptive. We expect it could be extended into a foundational technology for improving the reliability and adaptability of systems in which multiple agents must understand one another's states and intentions and work together, including multi-robot systems, autonomous drones, and human-AI collaboration."

Research Figure: MARBO's Overall Training Stages.

The findings were officially presented at EMNLP, an international conference on natural language processing.

This content was produced with the assistance of AI translation services.