iADD Optimizes Alignment and Diversity Tradeoffs in Diffusion Policies
September 30, 2026
iADD improves reinforcement learning-based post-training for diffusion models by addressing the diversity loss typically seen in reward optimization. The method uses incremental Feynman-Kac training to mathematically balance policy alignment with sample variety.
HOW THIS AFFECTS YOU
●
researcherThis provides a more stable theoretical framework for optimizing diffusion models via reward functions.