The Big Picture
Higher reported similarity between AI agents usually makes them act more cooperatively, but the effect varies widely by model and depends critically on how similarity is measured and shared.
ON THIS PAGE
Key Findings
Providing a numeric similarity score between two large language model agents often increases cooperation in classic social dilemmas. The strength of the effect differs a lot across models and game types, and cooperation shrinks when more players join or when the prompt frames agents as different. Grounding similarity in agents' observed decisions and explanations Semantic Capability Matching Pattern, letting agents judge each other, produces more consistent cooperation than arbitrary or mismatched benchmarks, yet some models still accept random similarity signals.
Test your agentsValidate against real scenarios
Key Data
19 LLMs tested across 10 grounding domains (7 public benchmarks + 3 custom domains).
2Theoretical thresholds: a homogeneous similarity above 50% supports full cooperation in Prisoner’s Dilemma and Public Good style games; >66% needed for the Travelers dilemma.
3Decision sampling used 10 runs per condition to estimate mean cooperation rates; cooperation patterns rose and fell dramatically across models and setups.
Why It Matters
Engineers designing systems of interacting agents and product leaders responsible for agent governance should care: similarity signals can be a lightweight lever to nudge agents toward mutually beneficial behavior. Researchers and safety teams should also care because the same signals can enable collusion or be gamed if not properly grounded and audited.
Key Figures
Fig 1: Figure 1: Overview of our evaluation framework. Left: Two LLM agents play a strategic game, such as the Prisoner’s Dilemma, and typically defect. Right: In our setup, agents are provided with a similarity score between them, computed based on their (individual) behavior on a chosen domain (“evaluation benchmark”). Cooperation emerges as the similarity score rises.
Fig 2: Figure 2: Cooperation rate in Prisoners for each of the 9 models, as a function of the reported similarity score. The ‘?’ and Base rows, respectively, show settings with an unspecified score or no mention of similarity.
Fig 3: Figure 3: Cooperation rates in Prisoners when similarity is grounded in any of 10 benchmarks, or not grounded (“No Domain”). For exact numbers, see Figure 12 in the appendix.
Fig 4: Figure 4: Pairwise similarity scores between five LLMs across four benchmarks, computed two ways: exogenously (top), and endogenously (bottom) by a judging model (rows) rating a target model (columns) given access to its decisions and explanations.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreConsiderations
Effects vary widely by model and by the exact cooperative game, so results don't generalize to all agent populations or settings. Some models accepted random or irrelevant similarity signals, exposing a risk that poorly grounded signals could be manipulated. Experiments focused on one-shot symmetric games and selected benchmarks; sequential or adaptive interactions may behave differently. Defense in Depth Pattern
Deep Dive
Experiments ran nine modern large language models on several classic cooperation problems (including Prisoner’s Dilemma, Public Good, and Travelers dilemmas) while giving agents a scalar similarity score between 0% and 100%. The study first treated similarity as an abstract signal and found that higher reported similarity usually raised cooperation rates, though the response curve and sensitivity differed dramatically by model. Variations in payoff scale, number of players, and prompt framing (emphasizing similarity vs difference) shifted cooperation predictably in many cases. Next, the authors grounded similarity by computing pairwise scores from agents’ behavior on ten benchmarks (moral, cognitive, personality-style tasks, plus three custom domains) and compared exogenous scores (computed by researchers) with endogenous scores (computed by agents judging each other’s decisions and explanations). Endogenous judgments—when agents saw peers’ decisions and explanations—yielded more consistent downstream cooperation than arbitrary or mismatched groundings, but some models still cooperated under randomly generated similarity values. LLM-as-Judge Pattern
To explain patterns, a simple behavioral model was introduced where an agent assumes that other agents will copy its contemplated deviation with a probability equal to the reported similarity. This model interpolates between classic independent decision-making (treat others as uncorrelated) and fully correlated reasoning where identical agents select the welfare-maximizing action. The theory gives concrete thresholds (e.g., >50% similarity suffices for some dilemmas) and bounds on welfare loss as similarity falls short of 100%. Practical implications: similarity signaling is a promising, low-cost mechanism to encourage cooperation among AI agents, but it carries reliability and collusion risks and must be grounded, monitored, and combined with safeguards before deployment. Rogue Agent Behavior
Explore evaluation patternsSee how to apply these findings
Credibility Assessment:
Contains a well-known researcher (Vincent Conitzer) and several other authors with modest h-indices; although arXiv and no citations, strong author reputation boosts credibility.