A new training method adds 24 points to AI win rates in games

The team of Professor Han Seung-yeol at UNIST in South Korea made MARBO. MARBO is a method to train language models, the AI that reads and writes text. It teaches an AI to find the roles of other AIs from their actions. It must not trust only their words. In the Werewolf game, the win rate increased from 41.6% to 65.6%.

How it works

First, the AI looks at the conversations and actions in the game. It uses them to find the role of each opponent and who is on its side. The team gives the name relational beliefs to these guesses.

The AI gets a reward only if its guess is correct and its actions help its goal. Other methods also gave a reward for an incorrect guess that caused a good result. MARBO does not count these.

The AI also checks if the opinion of the listener changes and this helps the AI.

The numbers

The team used 2 Mafia-style games, Werewolf and Among Us. In these games, players have roles that other players cannot see.

In Werewolf, the MARBO model won more frequently than a model with 2 times more parameters. That model had 64.0%. In Among Us, the win rate increased to 49.0% and 59.0% for 2 models with 2 and 4 billion parameters.

What is not known

The tests used games with AI opponents only. Han names systems of robots or drones as possible uses. The team gave no test of this.

Insight: Trust in agents moves from words to actions

MARBO trains an AI to check the actions of other AIs. Resemble AI says its detector tells a person, a bot and an AI agent apart. It uses cursor movement, typing rhythm and page steps. The two seem to trust actions more than words. Look for trust rules that include checks of actions. For builders, this is a sign to record actions.