English

Taming Preference Mode Collapse via Directional Decoupling Alignment in Diffusion Reinforcement Learning

Computer Vision and Pattern Recognition 2026-03-11 v2

Abstract

Recent studies have demonstrated significant progress in aligning text-to-image diffusion models with human preference via Reinforcement Learning from Human Feedback. However, while existing methods achieve high scores on automated reward metrics, they often lead to Preference Mode Collapse (PMC)-a specific form of reward hacking where models converge on narrow, high-scoring outputs (e.g., images with monolithic styles or pervasive overexposure), severely degrading generative diversity. In this work, we introduce and quantify this phenomenon, proposing DivGenBench, a novel benchmark designed to measure the extent of PMC. We posit that this collapse is driven by over-optimization along the reward model's inherent biases. Building on this analysis, we propose Directional Decoupling Alignment (D2^2-Align), a novel framework that mitigates PMC by directionally correcting the reward signal. Specifically, our method first learns a directional correction within the reward model's embedding space while keeping the model frozen. This correction is then applied to the reward signal during the optimization process, preventing the model from collapsing into specific modes and thereby maintaining diversity. Our comprehensive evaluation, combining qualitative analysis with quantitative metrics for both quality and diversity, reveals that D2^2-Align achieves superior alignment with human preference.

Keywords

Cite

@article{arxiv.2512.24146,
  title  = {Taming Preference Mode Collapse via Directional Decoupling Alignment in Diffusion Reinforcement Learning},
  author = {Chubin Chen and Sujie Hu and Jiashu Zhu and Meiqi Wu and Jintao Chen and Yanxun Li and Nisha Huang and Chengyu Fang and Jiahong Wu and Xiangxiang Chu and Xiu Li},
  journal= {arXiv preprint arXiv:2512.24146},
  year   = {2026}
}

Comments

Accepted by CVPR 2026

R2 v1 2026-07-01T08:45:38.961Z