RuleSmith: Multi-Agent LLMs for Automated Game Balancing
RuleSmith automates game balancing by using multi-agent LLMs for self-play and Bayesian optimization over a rule space. It tests CivMini, a simplified strategy game with two asymmetric factions, to find rule configurations that result in near-even win rates. The framework uses adaptive sampling to efficiently search for balanced parameters without requiring human expert tuning.