Key Takeaway
Privately asking each agent about what others do and what should be done raises cooperation and reveals whether group behavior rests on shared beliefs or simple imitation.
ON THIS PAGE
The Evidence
**WHAT THEY FOUND:** Prompting agents to report their expectations consistently increased contributions across four model families and made agents form shared beliefs about typical behavior shared beliefs about typical behavior. Reports about what others actually do (empirical expectations) tended to converge faster and more tightly than beliefs about what ought to be done (normative expectations). Social learning helped stabilize cooperation, social selection helped identify cooperators but had weaker direct effects on contributions, and after adversarial disruption expectations bounced back faster than actual behavior. adversarial disruption
Explore evaluation patternsSee how to apply these findings
Data Highlights
1Expectation elicitation was the only mechanism that improved average contributions across all four model families (results significant in OLS; 10 independent runs per model-condition).
2A larger gap between an agent's contribution and its normative expectation predicted next-round adjustment with β = 0.210, while the empirical expectation gap predicted adjustment with β = 0.169 (NE effect significantly stronger, p = .038).
3GPT-family agents returned null normative reports 25% of the time, compared with under 3% nulls for the other model families; experiments used n = 10 runs per model-condition.
Why It Matters
**WHO SHOULD CARE:** Engineers building multi-agent systems who need tools to diagnose whether group cooperation is robust or superficial; they can use expectation elicitation as a lightweight signal to detect and nudge coordination. Technical leads and researchers evaluating agent governance and agent-to-agent evaluation can use these measurements to separate mere behavioral convergence from true shared beliefs that sustain cooperation. For practical guidance, see governance-oriented patterns Guardrails Pattern.
Key Figures
Fig 1: Figure 1: Our framework for mechanism ablation of social norm emergence in generative agent societies. The agents play a public good games in groups where they can discuss and create expectations about appropriate behavior.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreLimitations
**CONSIDERATIONS:** The setting uses a simplified public goods game, so results may not generalize to all real-world multi-agent tasks. Expectation reports come from prompt responses, not direct access to internal model states, so they are indicative rather than conclusive evidence of internal beliefs. Experiments cover four model families and modest run counts, so effects should be validated in larger, more diverse deployments before production use. For real-world deployment considerations, see applicable use-case insights Multi-Agent E-Commerce Operations
The Details
**FULL SUMMARY:** Simulations ran 12 agents split into groups of four over 20 rounds of a public goods game where each agent chose how much to contribute from an endowment. After each round agents could discuss, choose contributions, evaluate peers, and were privately asked two questions: what they expected others to contribute (empirical expectation) and what they thought others ought to contribute (normative expectation). The study ablated two group mechanisms—social learning (how behavior spreads) and social selection (partner choice and reputation)—and added an adversarial disruption to test recovery dynamics. Eliciting expectations raised contributions across all tested models and led to tighter agreement about what others actually do than about what others should do. Both kinds of expectation gaps predicted how agents adjusted their contributions next round (normative gap β = 0.210, empirical gap β = 0.169), with normative expectations having a stronger directional effect overall. Social learning made contributions more stable; social selection helped pick out cooperators but did not strongly increase contributions by itself. After adversarial disturbance, agents' expectation reports recovered faster than their actions, showing that shared beliefs can re-emerge before behavior fully follows. Practical takeaway: adding expectation elicitation is a low-cost diagnostic (and possible coordination) tool for multi-agent trust and agent-to-agent evaluation, but it should be paired with broader mechanism tests and real-world validation. For broader methodology guidance, consider defense-in-depth practices Defense in Depth Pattern and synthesis across patterns Multi-Agent Research Synthesis.
Need expert guidance?We can help implement this
Credibility Assessment:
ArXiv preprint but includes an author with moderate h-index (h=16) and another with h=6; suggests a recognized/solid researcher mix despite no venue listed.