Talk, Judge, Cooperate: Gossip-Driven Indirect Reciprocity in Self-Interested LLM Agents

Researchers created a system called ALIGN where AI agents share gossip about each other's behavior to decide who to trust. This decentralized 'reputation' system helps self-interested agents cooperate more effectively by identifying and ostracizing bad actors. The study found that AI models with stronger reasoning skills cooperated more strategically, while simpler chat models often cooperated too much, even when it was disadvantageous. This approach improves group welfare without changing the agents' core programming.