Guardado en:
| Autores principales: | , , , , , , , , , , , , , , , , , |
|---|---|
| Formato: | Preprint |
| Publicado: |
2025
|
| Materias: | |
| Acceso en línea: | https://arxiv.org/abs/2512.04332 |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| _version_ | 1866912787394985984 |
|---|---|
| author | Ye, Haotian Zheng, Kaiwen Xu, Jiashu Li, Puheng Chen, Huayu Han, Jiaqi Liu, Sheng Zhang, Qinsheng Mao, Hanzi Hao, Zekun Chattopadhyay, Prithvijit Yang, Dinghao Feng, Liang Liao, Maosheng Bai, Junjie Liu, Ming-Yu Zou, James Ermon, Stefano |
| author_facet | Ye, Haotian Zheng, Kaiwen Xu, Jiashu Li, Puheng Chen, Huayu Han, Jiaqi Liu, Sheng Zhang, Qinsheng Mao, Hanzi Hao, Zekun Chattopadhyay, Prithvijit Yang, Dinghao Feng, Liang Liao, Maosheng Bai, Junjie Liu, Ming-Yu Zou, James Ermon, Stefano |
| contents | Aligning generative diffusion models with human preferences via reinforcement learning (RL) is critical yet challenging. Most existing algorithms are often vulnerable to reward hacking, such as quality degradation, over-stylization, or reduced diversity. Our analysis demonstrates that this can be attributed to the inherent limitations of their regularization, which provides unreliable penalties. We introduce Data-regularized Diffusion Reinforcement Learning (DDRL), a novel framework that uses the forward KL divergence to anchor the policy to an off-policy data distribution. Theoretically, DDRL enables robust, unbiased integration of RL with standard diffusion training. Empirically, this translates into a simple yet effective algorithm that combines reward maximization with diffusion loss minimization. With over a million GPU hours of experiments and ten thousand double-blind human evaluations, we demonstrate on high-resolution video generation tasks that DDRL significantly improves rewards while alleviating the reward hacking seen in baselines, achieving the highest human preference and establishing a robust and scalable paradigm for diffusion post-training. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2512_04332 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Data-regularized Reinforcement Learning for Diffusion Models at Scale Ye, Haotian Zheng, Kaiwen Xu, Jiashu Li, Puheng Chen, Huayu Han, Jiaqi Liu, Sheng Zhang, Qinsheng Mao, Hanzi Hao, Zekun Chattopadhyay, Prithvijit Yang, Dinghao Feng, Liang Liao, Maosheng Bai, Junjie Liu, Ming-Yu Zou, James Ermon, Stefano Machine Learning Aligning generative diffusion models with human preferences via reinforcement learning (RL) is critical yet challenging. Most existing algorithms are often vulnerable to reward hacking, such as quality degradation, over-stylization, or reduced diversity. Our analysis demonstrates that this can be attributed to the inherent limitations of their regularization, which provides unreliable penalties. We introduce Data-regularized Diffusion Reinforcement Learning (DDRL), a novel framework that uses the forward KL divergence to anchor the policy to an off-policy data distribution. Theoretically, DDRL enables robust, unbiased integration of RL with standard diffusion training. Empirically, this translates into a simple yet effective algorithm that combines reward maximization with diffusion loss minimization. With over a million GPU hours of experiments and ten thousand double-blind human evaluations, we demonstrate on high-resolution video generation tasks that DDRL significantly improves rewards while alleviating the reward hacking seen in baselines, achieving the highest human preference and establishing a robust and scalable paradigm for diffusion post-training. |
| title | Data-regularized Reinforcement Learning for Diffusion Models at Scale |
| topic | Machine Learning |
| url | https://arxiv.org/abs/2512.04332 |