Mixture-Greedy Strategy for Generative Model Selection

The Mixture-Greedy approach outperforms traditional Upper Confidence Bound (UCB) methods in diversity-aware multi-armed bandit tasks for model selection. It achieves faster convergence and better performance on metrics like FID and Vendi by avoiding explicit optimism terms that slow down sampling.