Abstract
As AI agents become increasingly integrated into human social structures, they must navigate complex environments characterized by conflicting interests and power asymmetries. This paper introduces the Junior High Game (JHG), a novel social test-bed that simulates these dynamics. We present Community-Aware Behavior (CAB), an agent architecture that employs network science to form and maintain cooperative alliances. Our experiments, involving both human-bot studies and large-scale simulations, demonstrate that CAB agents can achieve human-level performance and enhance overall social welfare. However, we also identify a critical "policing" vulnerability, where agents struggle to defend their communities against coordinated adversarial attacks.
Introduction: Beyond Zero-Sum Coordination
Traditional AI benchmarks often focus on zero-sum competition (e.g., Chess, Go) or simple coordination (e.g., Overcooked). Real-world human societies are significantly more complex, featuring non-stationarity, power imbalances, and intricate interdependencies. To address this, we propose a framework for evaluating AI's ability to engage in collective action—coordinating with others toward shared goals in "general-sum" environments where individual and group interests may clash.
The Junior High Game (JHG)
The JHG is a model of social systems where players compete for "popularity." In each round, players allocate tokens to three actions:
- Keep: Increases the player's own popularity and defensive strength.
- Give: Increases another player's popularity (a positive-sum trade).
- Take: Attacks another player's popularity (a negative-sum conflict).
Crucially, a player's current popularity acts as a multiplier for their actions, creating a feedback loop where the "popular kids" wield more influence. This mechanism naturally leads to power asymmetries and the formation of shifting alliances.
Community-Aware Behavior (CAB)
To navigate the JHG, we developed the CAB agent. CAB operates in three main steps:
- Detect Communities: Using the Louvain Method to identify cliques based on historical token exchanges.
- Select a Community: Evaluating potential groups based on factors like group strength, familiarity, and prosocial behavior.
- Allocate Tokens: Distributing resources to support the chosen community or retaliate against perceived threats.
We optimized CAB parameters using evolutionary algorithms across different initial wealth classes (poor, middle class, and rich) to ensure robust performance across social strata.
Experimental Results and Social Impact
Our evaluation included human-bot user studies with 24 participants and extensive simulations to test agent resilience.
Human-Level Performance
In mixed-player games, CAB agents achieved popularity levels statistically indistinguishable from experienced human players ($p = 0.132$). While humans were significantly more aggressive (attacking 8.4% of the time vs. CAB's 0.08%), CAB agents focused on reciprocal "giving" to build stable economic alliances. CAB agents also tended to keep more tokens for themselves (37% vs. 16% for humans), prioritizing individual stability alongside group cooperation.
Social Welfare and Inequality
A key finding is that societies with a higher density of CAB agents exhibit higher Social Welfare (average popularity) and lower Gini Indices (wealth inequality). CAB's preference for positive-sum "giving" over destructive "taking" acts as a stabilizing force, preventing the "popularity wars" often seen in human-only groups.
The "Policing" Problem and Adversarial Threats
Despite their success in building communities, CAB agents showed a marked vulnerability to CATs (Conspiring Autonomous Thieves)—adversarial agents that coordinate to pillage others. Standard evolved CAB agents failed to recognize when allies were under attack and did not coordinate counter-strikes. This "policing" failure suggests that while community detection is a powerful tool for cooperation, defending those communities requires explicit, higher-order reasoning about adversarial intent. When retaliation logic was manually added, CAB agents successfully suppressed bullies, highlighting the need for proactive defense mechanisms.
Conclusion
The Junior High Game provides a rich environment for studying social AI. Our work shows that community-aware strategies can foster collective action and improve social outcomes. Future research will focus on developing dynamic learning mechanisms that allow agents to adapt to shifting social winds and more effectively mitigate coordinated threats in larger, more complex societies.
Summary updated