LLM-Judged Open-Strategy Dictator Game: Conditional Cooperation Wins

A new arXiv paper from researcher Michael Glass proposes the Open-Strategy Dictator Game (OSDG), a twist on the classic dictator game where each player's strategy is a natural-language document fully visible to all participants. The dictator's decision to SHARE or TAKE an endowment can depend on the recipient's written strategy, and a large language model adjudicates each interaction by interpreting the dictator's strategy in the context of the recipient's. The setup models real-world multi-agent systems where AI agents can inspect each other's decision procedures.

Glass ran round-robin tournaments among diverse strategies and analyzed the resulting payoff matrix using softmax equilibrium frequencies, dominance analysis, and sensitivity to the relative value of cooperation. The results are clear: conditionally cooperative strategies—those that share with cooperators and take from exploiters—consistently dominate across a wide range of payoff parameters. Unconditional strategies (always share or always take) are weakly dominated. This suggests that in environments where agents can observe each other's decision rules, reciprocal cooperation is not just nice—it's evolutionarily stable. For professionals building multi-agent AI systems, the implication is significant: transparent, rule-based cooperation policies could be the key to designing resilient and trustworthy agent swarms.