Researchers propose incentive mechanisms to keep AI agents in checkMachine translation

In a preprint, MIT economist Stephen Morris and colleagues propose rewarding a model for how well its answers predict another model’s responses, making deliberate underperformance on tests less attractive. Princeton’s Elad Hazan envisions models taking turns solving problems and auditing each other, with rewards and penalties tied to catching cheating. The approaches remain untested on actual frontier models and could face collusion between agents.