Beyond the Tragedy of the Commons: Building A Reputation System for Generative Multi-agent Systems

TL;DR

This work tackles the tragedy of the commons in generative multi-agent systems by introducing RepuNet, a dynamic reputation framework that couples agent-level reputation dynamics with system-level network evolution driven by direct encounters and gossip. By shaping both self- and peer-reputations through encounters and gossip, and by rewiring networks to favor reputable partners, RepuNet incentivizes cooperative behavior and prevents resource exploitation. Two scenarios, a voluntary participation public goods task and a trust-based trading game, demonstrate that RepuNet fosters cooperative clusters, isolates defectors, and even reveals a bias toward sharing positive gossip, with an ablation study confirming the necessity of each component. The findings suggest reputation-driven social signaling and network adaptation as scalable mechanisms to maintain cooperation in generative MASs, with implications for robust collaboration in AI ecosystems.

Abstract

The tragedy of the commons, where individual self-interest leads to collectively disastrous outcomes, is a pervasive challenge in human society. Recent studies have demonstrated that similar phenomena can arise in generative multi-agent systems (MASs). To address this challenge, this paper explores the use of 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 two distinct scenarios, we show that RepuNet effectively mitigates the 'tragedy of the commons', promoting and sustaining cooperation in generative MASs. Moreover, we find that reputation systems can give rise to rich emergent behaviors in generative 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.

Figures (15)

• Figure 1: RepuNet: A dynamic reputation system aimed at sustaining cooperation and preventing the 'tragedy of the commons' in generative MASs. Agents interact within networks through both direct encounters and indirect gossip, shaping their self-reputation and peer-reputation. At the system level, agents decide whether to form or maintain network connections (i.e. edges) based on these reputations. The evolving reputations and network structures, stored in RepuNet databases, continuously guide agents' behaviors, influencing their future interactions.
• Figure 2: Experimental results in both scenarios: For Scenario 1, panel (a) depicts agent participation trends, with solid lines showing participation rates and the shaded area indicating the error margin. For Scenario 2, panel (c) depicts investment success trends, with solid lines showing success rates and the shaded area indicating the error margin. Panels (b) and (d) both show the statistically significant ($p<0.002$) correlation between agent behavior and reputation, highlighted by the linear regression line. Different shades of blue in panels (b) and (d) represent five experimental runs, with each color group showing the average behavior of 20 agents over the last ten rounds.
• Figure 3: A case study on network dynamics at the system level. Nodes represent agents, with colors from blue (low reputation) to red (high reputation); green indicates agents without RepuNet. Node size reflects bidirectional connections. In Scenario 1, panel (a) and (b) show that agents without RepuNet connect indiscriminately, while those with RepuNet form dense clusters with reputable partners and isolate low-reputation agents. In Scenario 2, networks without RepuNet collapse due to defective behavior (the 'tragedy of the commons'), whereas networks with RepuNet gradually form clusters of reputable agents, ultimately enabling all agents to achieve high reputation and cluster together.
• Figure 4: Correlation between gossip frequency and reputation in Scenario 1. Agents who are discussed more frequently in gossip tend to earn higher reputations, supported by our sentiment analysis showing 90% of gossip is positive (e.g., praise for cooperation). The linear regression line highlights the statistically significant positive trend ($p<0.002$). Different shades of blue represent five experimental runs, with each color group showing the average behavior of 20 agents.
• Figure S1: Prompt for $r_{i\rightarrow j}(t+1) \leftarrow \texttt{ShapeRepuPeer}(m_{ij}(t+1),\text{if}\ r_{i\rightarrow j}(t)=\phi)$: Reputation driven by direct encounters at the agent level.