GARDEN reward design lifts unseen-scenario win rate to 89.9% in cooperative MARL

[arXiv] A paper titled GARDEN (Generalist Autonomous Reward Design through Evidence Navigation), by Dogan Urgun, proposes a method for cooperative multi-agent reinforcement learning in which a large language model acts as a reward engineer, examining environment code, training records and agent behavior and testing draft rewards in trial training runs while the learning algorithm and the rule for selecting the final reward are held fixed. The approach develops one reward program for a family of related scenarios and transfers it unchanged to scenarios unavailable during design. According to the paper, on unseen scenarios of the Overcooked and SMAX benchmarks, task returns rose 58.2% and 52.5% over default rewards on two kitchen layouts, and mean win rate in two combat scenarios increased from 51.0% to 89.9%. Under matched candidate and probe budgets, the paper reports this exceeded the 84.2% obtained with fixed diagnostic feedback and 79.2% with Eureka-inspired reward reflection. The authors say the results show a single autonomously engineered reward can be reused effectively across related cooperative scenarios. The work is a preprint and has not been peer reviewed.