Do LLMs Take Care of Their Own? Similarity Signals Can Induce Cooperation

2026-08-12

AI Summary

This study addresses the challenge that large language model (LLM) agents face in achieving mutually beneficial cooperation in strategic interactions when explicit similarity cues are absent. The work proposes the first systematic evaluation framework to investigate LLM cooperation under varying similarity signals, integrating game-theoretic experiments, chain-of-thought reasoning, and equilibrium modeling with quantifiable similarity indicators. The findings reveal that certain modern LLMs consistently cooperate across multiple games in response to similarity signals, with cooperation propensity driven more significantly by signal strength than by data provenance. Moreover, LLMs generally overestimate their similarity to other agents and converge to cooperative equilibria when perceiving high similarity. This research establishes a behavioral game-theoretic foundation for understanding and fostering cooperation among LLM agents.

Abstract

As LLM-based agents with user-instructed goals are becoming widely deployed, they increasingly encounter each other in strategic interactions, and face challenges of finding mutually beneficial outcomes. Prior literature has argued that cooperation problems such as the Prisoner's Dilemma are resolvable in settings where agents know they follow very similar decision making patterns, as for example in monocultural AI ecosystems. Following that line of work, this paper introduces the first framework for evaluating LLM decision making when agents are provided with graded similarity signals. Among our findings, we establish 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. Perhaps surprisingly, our experiments also show that the dataset based on which the similarity signal is computed has small to no impact on induced cooperation, and that LLM models systematically self-identify as highly similar when asked to evaluate another model's chain-of-thought reasoning by themselves. Finally, we develop an LLM-behavioral-game-theoretic model that captures some of their reasoning rationale, and show that it can support cooperative outcomes in equilibrium under sufficiently high similarity scores.

Problem

• LLM cooperation
• similarity signals
• strategic interaction
• Prisoner's Dilemma
• AI alignment

Innovation

• similarity signals
• LLM cooperation
• behavioral game theory
• chain-of-thought reasoning
• multi-agent interaction