Do LLMs Take Care of Their Own? Similarity Signals Can Induce Cooperation
A paper titled 'Do LLMs Take Care of Their Own? Similarity Signals Can Induce Cooperation' was published on arXiv (cs.AI) on 2026-08-12. It introduces a framework for evaluating LLM decision making when agents are provided with graded similarity signals. The paper reports that different LLM models vary drastically in how they navigate similarity signals, with some modern models showing consistent behavior across cooperation problems, payoff structures, and prompt framing. It also finds that the dataset used to compute similarity signals has small to no impact on induced cooperation, and that LLM models systematically self-identify as highly similar when evaluating another model's chain-of-thought reasoning. The paper develops an LLM-behavioral-game-theoretic model.
Researchers introduced the first framework for evaluating LLM decision making when agents receive graded similarity signals, focusing on cooperation problems like the Prisoner's Dilemma. They found that LLM models differ significantly in their responses to similarity signals, with some modern models behaving consistently across various cooperation problems, payoff structures, and prompt framings. The dataset used to compute similarity had minimal impact on induced cooperation, and LLMs tended to self-identify as highly similar when assessing another model's chain-of-thought reasoning. The work also includes an LLM-behavioral-game-theoretic model.
The framework evaluates LLM decision making under graded similarity signals, revealing model-specific behavioral patterns. The finding that dataset choice has little effect on cooperation suggests that similarity signals may be processed in a way that is robust to the underlying data distribution. The self-identification of high similarity in chain-of-thought evaluations indicates a potential bias in LLM self-assessment. The developed LLM-behavioral-game-theoretic model could provide a formal basis for predicting cooperative outcomes in multi-agent LLM systems.
As LLM-based agents are increasingly deployed in strategic interactions, understanding how similarity signals influence cooperation is relevant for designing multi-agent systems, marketplaces, and negotiation platforms. The observed variability across models suggests that deployment choices may affect cooperative outcomes. The minimal impact of dataset choice could simplify implementation, as less effort may be needed to curate similarity data. The self-identification bias may have implications for trust and verification in agent interactions.
For businesses deploying LLM agents in competitive or cooperative settings, this research suggests that model selection and similarity signal design can influence cooperative outcomes. The finding that dataset choice has minimal impact may reduce data engineering costs. Understanding self-identification bias could help in building more reliable agent evaluation systems. The framework could be used to test and improve agent behavior in negotiation, resource allocation, and collaborative tasks.
Next signals to watch include follow-up studies testing the framework on additional LLM models and real-world multi-agent scenarios, as well as applications of the behavioral-game-theoretic model to agent design. Potential developments include standardized benchmarks for cooperation in LLM agents and integration of similarity signals into agent communication protocols. The research may also prompt investigations into mitigating self-identification bias in chain-of-thought evaluations.