Study: LLM agents cooperate more when similarity signals are strong

As LLM-based agents increasingly negotiate with each other in real-world applications, understanding when they cooperate or defect is critical. In a new arXiv paper, "Do LLMs Take Care of Their Own? Similarity Signals Can Induce Cooperation," Akash Kundu and colleagues from Carnegie Mellon University and other institutions present the first framework to systematically test LLM behavior in strategic games when agents receive graded signals about how similar their decision-making processes are. The work builds on prior theory suggesting that cooperation in dilemmas like the Prisoner's Dilemma becomes attainable when agents believe they share similar reasoning patterns, as in monocultural AI ecosystems.

Across experiments with multiple modern LLMs, the authors found dramatic differences in how models react to similarity cues. Some models exhibited stable cooperation across varied payoff structures and prompt framings, while others shifted strategies unpredictably. Perhaps the most surprising result: the specific dataset used to compute the similarity signal had negligible effect on cooperation rates. Furthermore, when asked to evaluate another model's chain-of-thought reasoning, LLMs systematically rated it as highly similar to their own. The researchers also developed a behavioral game-theoretic model that captures these dynamics, showing that cooperative equilibria emerge only when similarity scores exceed a threshold. These findings have practical implications for designing multi-agent AI systems—from automated negotiation to decentralized coordination—where encouraging perceived alignment could be a lever for safer, more collaborative outcomes.