Reputation as a Solution to Cooperation Collapse in LLM-based MASs

Siyue Ren, Wanli Fu, Xinkun Zou, Chen Shen, Yi Cai, Chu Chen, Zhen Wang, Shanshan Hu

Abstract

Cooperation has long been a fundamental topic in both human society and AI systems. However, recent studies indicate that the collapse of cooperation may emerge in multi-agent systems (MASs) driven by large language models (LLMs). To address this challenge, we explore reputation systems as a remedy. We propose RepuNet, a dynamic, dual-level reputation framework that models both agent-level reputation dynamics and system-level network evolution. Specifically, driven by direct interactions and indirect gossip, agents form reputations for both themselves and their peers, and decide whether to connect or disconnect other agents for future interactions. Through three distinct scenarios, we show that RepuNet effectively avoids cooperation collapse, promoting and sustaining cooperation in LLM-based MASs. Moreover, we find that reputation systems can give rise to rich emergent behaviors in LLM-based MASs, such as the formation of cooperative clusters, the social isolation of exploitative agents, and the preference for sharing positive gossip rather than negative ones.

Introduction

As Large Language Model (LLM) agents transition from isolated assistants to complex Multi-Agent Systems (MASs), they encounter a classic sociological hurdle: cooperation collapse. In environments where individual gains conflict with collective welfare, LLM agents often fall into "death loops" of mutual defection, triggered by short-term greed or over-retaliation to single errors. Without a mechanism for social accountability, these digital societies lack the trust necessary for sustained utility.

To bridge this gap, we introduce RepuNet, a framework designed to foster long-term cooperation by enabling agents to track reputations and dynamically restructure their social networks.

The RepuNet Framework

RepuNet operates through a dual-level architecture that mirrors human social dynamics, utilizing natural language prompting to allow LLMs to reason about social standing rather than relying on hard-coded formulas.

Overview of the RepuNet framework, illustrating the dual-level dynamics of agent reputation and system-level network evolution driven by direct encounters and indirect gossip.

1. Agent-Level Reputation Dynamics

Following an interaction, agents update their internal records in a "RepuNet Database." This includes self-reputation (an agent's perception of how others view it) and peer-reputation (its evaluation of a partner's trustworthiness).

2. System-Level Network Evolution

Reputation directly influences the system's topology. Agents autonomously decide to build new edges with reputable peers or sever ties with known defectors, effectively "voting with their feet" to maintain a healthy social environment.

3. Information Channels: Direct vs. Indirect

Reputation is built through direct experience and an indirect gossip mechanism. Agents share information about third parties, and listeners evaluate the credibility of this gossip based on the gossiper's own reputation before updating their records.

Experimental Scenarios

We evaluated RepuNet using 20 LLM agents (primarily GPT-4o mini, with validation on Qwen and Gemini) across three social dilemmas:

  • The Prisoner’s Dilemma: Testing fundamental cooperation vs. defection.
  • Voluntary Participation: A public goods game modeled on energy-saving programs.
  • Trading Investment: A trust game testing sequential honesty and resource allocation.

Key Results and Emergent Behaviors

Our experiments demonstrate that RepuNet is a robust solution to cooperation collapse. In the Prisoner's Dilemma, agents using RepuNet maintained a 93% cooperation rate, while the control group (without reputation) plummeted to 9%. Similarly, in the Trading Investment scenario, success rates reached 98% with RepuNet compared to total system disintegration without it.

(a) Comparison of cooperation rates over time with and without RepuNet. (b) Correlation between an agent's reputation and its probability of cooperation.

Beyond raw metrics, several rich social behaviors emerged:

  • Cooperative Clustering: High-reputation agents naturally formed dense, mutually beneficial networks, while defectors were socially isolated.
  • Positive Gossip Bias: Interestingly, LLM agents predominantly shared positive gossip (~90% of cases), preferring to praise good partners rather than complain about bad ones—a notable departure from typical human social patterns.
  • Model Robustness: The framework's success across different LLM families suggests that "reputation" is a high-level concept that modern models grasp intuitively.

Conclusion and Future Work

RepuNet proves that explicit reputation frameworks are essential for stable LLM societies; simple interaction memory is insufficient to prevent collapse. While highly effective, future research should address the potential for dishonest gossip (manipulative agents) and the scalability of these dynamics to systems involving thousands of agents. RepuNet represents a critical first step toward building resilient, self-governing AI communities.

Summary updated

Abstract

Cooperation has long been a fundamental topic in both human society and AI systems. However, recent studies indicate that the collapse of cooperation may emerge in multi-agent systems (MASs) driven by large language models (LLMs). To address this challenge, we explore reputation systems as a remedy. We propose RepuNet, a dynamic, dual-level reputation framework that models both agent-level reputation dynamics and system-level network evolution. Specifically, driven by direct interactions and indirect gossip, agents form reputations for both themselves and their peers, and decide whether to connect or disconnect other agents for future interactions. Through three distinct scenarios, we show that RepuNet effectively avoids cooperation collapse, promoting and sustaining cooperation in LLM-based MASs. Moreover, we find that reputation systems can give rise to rich emergent behaviors in LLM-based MASs, such as the formation of cooperative clusters, the social isolation of exploitative agents, and the preference for sharing positive gossip rather than negative ones. The GitHub repository for our project can be accessed via the following link: https://github.com/RGB-0000FF/RepuNet.
BibTeX
@misc{ren2025reputation,
  author = {Siyue Ren and Wanli Fu and Xinkun Zou and Chen Shen and Yi Cai and Chu Chen and Zhen Wang and Shanshan Hu},
  title = {{Reputation as a Solution to Cooperation Collapse in LLM-based MASs}},
  year = {2025},
  eprint = {2505.05029},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.05029},
}
Figure