CoRe Framework Mitigates Latent Reward Hacking in Video Diffusion
September 30, 2026
CoRe prevents video diffusion models from drifting into out-of-distribution latent spaces by co-evolving the reward model with the generator. This dynamic interaction prevents reward hacking where perceived quality drops even as latent reward scores increase.
HOW THIS AFFECTS YOU
●
builderYou can improve video generation stability by refitting reward models on current generator samples.
●
researcherThis addresses the distributional escape problem in latent reward model optimization.