New benchmark reveals stronger LLM agents defect more in cooperative games, spurring research into game-theoretic safeguards

Recent LLM models consistently defect in single-shot social dilemmas like prisoner's dilemma and public goods games, with stronger reasoning capabilities paradoxically reducing cooperative behavior

CoopEval introduces the first comparative study evaluating four cooperation-enabling mechanisms: repeated gameplay, reputation systems, third-party mediators, and contract-based agreements

Research addresses critical safety concern: ensuring LLM agents can interact effectively and safely with other goal-pursuing agents in mixed-motive game scenarios

Study tests mechanisms across four distinct social dilemmas to identify which game-theoretic approaches enable rational agents to reach cooperative equilibrium outcomes