@inproceedings{chen-etal-2026-apex,
    title = "{APEX}: Learning Adaptive Priorities for Multi-Objective Alignment in Vision-Language Generation",
    author = "Chen, Dongliang  and
      Zhuang, Xinlin  and
      Xu, Junjie  and
      Xie, Luojian  and
      Wang, Zehui  and
      Zhuang, Jiaxi  and
      Yang, Haolin  and
      Dou, Liang  and
      He, Xiao  and
      Wu, Xingjiao  and
      Qian, Ying",
    editor = "Liakata, Maria  and
      Moreira, Viviane P.  and
      Zhang, Jiajun  and
      Jurgens, David",
    booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {ACL} 2026",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-acl/2026.findings-acl.243/",
    pages = "4922--4939",
    ISBN = "979-8-89176-395-1",
    abstract = "Multi-objective alignment for text-to-image generation is commonly implemented via static linear scalarization, but \textit{fixed} weights often fail under heterogeneous rewards, leading to optimization imbalance where models overfit high-variance, high-responsiveness objectives (e.g., OCR) while under-optimizing perceptual goals. We identify two mechanistic causes: \textbf{variance hijacking}, where reward dispersion induces implicit reweighting that dominates the normalized training signal, and \textbf{gradient conflicts}, where competing objectives produce opposing update directions and trigger seesaw-like oscillations. We propose \textbf{APEX} (\textbf{A}daptive \textbf{P}riority-based \textbf{E}fficient \textbf{X}-objective Alignment), which stabilizes heterogeneous rewards with \textbf{Dual-Stage Adaptive Normalization} and dynamically schedules objectives via \textbf{$\mathcal{P}^3$ Adaptive Priorities} that combine learning potential, conflict penalty, and progress need. On Stable Diffusion 3.5, APEX achieves improved Pareto trade-offs across four heterogeneous objectives, with balanced gains of \textbf{+1.31 PickScore}, \textbf{+0.35 DeQA}, and \textbf{+0.53 Aesthetics} while maintaining competitive OCR accuracy, mitigating the instability of multi-objective alignment."
}