CoRe: Co-evolving reward models to mitigate latent reward hacking in video diffusion models

To address the problem that optimizing against a fixed latent reward model leads to latent reward hacking and degradation of perceptual and motion quality in generated videos, the researchers propose the co-evolving reward framework CoRe, which treats latent-space alignment as a dynamic interaction between the generator and the reward model, continuously refitting the reward model with the generator's current samples and anchoring to real video preferences, so that the generator cannot obtain high rewards by deviating from the data distribution. On Wan2.1-T2V-1.3B, CoRe's generation quality consistently outperforms the pretrained model and existing alignment methods, and avoids the quality collapse caused by fixed-reward optimization.