Sustaining indirect reciprocity among decentralized self-interested LLM agents

Determine whether and how self-interested large language model (LLM) agents can sustain indirect reciprocity in mixed-motive tasks without compromising the decentralized, self-interested setting.

Background

The paper studies cooperation among decentralized, self-interested LLM agents, where direct reciprocity is often unavailable due to large, dynamic populations and one-shot pairings. Indirect reciprocity requires reputation mechanisms that allow agents to judge others based on their behavior toward third parties.

Classical models (e.g., image scoring and second-order norms) typically assume centralized monitoring and static norms, which do not naturally transfer to decentralized LLM-agent systems. This raises a fundamental question about the viability of maintaining indirect reciprocity among self-interested LLM agents in mixed-motive environments without introducing centralized control or altruism.

References

Therefore, it is still an open question whether and how self-interested LLM agents can sustain indirect reciprocity in mixed-motive tasks without compromising the decentralized self-interested setting.

— Talk, Judge, Cooperate: Gossip-Driven Indirect Reciprocity in Self-Interested LLM Agents  (2602.07777 - Zhu et al., 8 Feb 2026) in Section 1 (Introduction)