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

TL;DR

The paper tackles the challenge of sustaining indirect reciprocity among decentralized, self-interested LLM agents. It introduces ALIGN, an automated framework that uses hierarchical, open-ended gossip to evaluate trust, coordinate social norms, and guide actions, enabling indirect reciprocity without central authority. Through game-theoretic analysis and extensive experiments across matrix games, investment scenarios, and a transaction market, the authors show that public gossip can sustain cooperation in infinite-horizon settings when future payoffs are sufficiently valued (gamma near or above c/b), and that stronger LLM reasoning generally leads to more incentive-aligned cooperation. Ablation and robustness analyses demonstrate that gossip is a critical driver for cooperation, while reflection and equilibrium knowledge modulate but do not replace the core benefits of gossip; the framework also shows resilience to untruthful signals and misreports. Together, these results suggest gossip-driven decentralized reputation mechanisms as a practical path to maintaining welfare in agentic ecosystems, with open questions about privacy, fairness, and potential manipulation in real-world deployments.

Abstract

Indirect reciprocity, which means helping those who help others, is difficult to sustain among decentralized, self-interested LLM agents without reliable reputation systems. We introduce Agentic Linguistic Gossip Network (ALIGN), an automated framework where agents strategically share open-ended gossip using hierarchical tones to evaluate trustworthiness and coordinate social norms. We demonstrate that ALIGN consistently improves indirect reciprocity and resists malicious entrants by identifying and ostracizing defectors without changing intrinsic incentives. Notably, we find that stronger reasoning capabilities in LLMs lead to more incentive-aligned cooperation, whereas chat models often over-cooperate even when strategically suboptimal. These results suggest that leveraging LLM reasoning through decentralized gossip is a promising path for maintaining social welfare in agentic ecosystems. Our code is available at https://github.com/shuhui-zhu/ALIGN.

Figures (15)

• Figure 1: Illustration of three monitoring structures: (a) Private monitoring, only the donor and recipient observe the donor's action; (b) Perfect public monitoring, all agents observe the donor's action; (c) Imperfect public monitoring via gossip, only the donor and recipient observe the action, and all agents observe the public signal broadcast by the recipient.
• Figure 2: Decision Process in ALIGN
• Figure 3: Generative Agent Architecture of ALIGN
• Figure 4: Tones of Gossip Protocol
• Figure 5: Tone Proportions in Repeated Donation Game: (a) cooperating chat models, (b) cooperating reasoning models, (c) defecting chat models and (d) defecting reasoning models. ALIGN agents typically praise cooperation and criticize defection.

Theorems & Definitions (10)

• Definition 3.1: Repeated Donation Game
• Proposition 3.2
• Proposition 3.3
• Proposition 3.4
• Proposition 3.5
• proof : Proof of Proposition \ref{['prop_fh']}
• proof : Proof of Proposition \ref{['prop_ifh_nomonitor']}
• proof : Proof of Proposition \ref{['prop_ifh_monitor']}
• Definition 1.1: Repeated Donation Game with Public Gossip
• proof : Proof of Proposition \ref{['prop_ifh_signal']}