面向生成式用户界面的高效个性化
星系天体物理
2026-04-16 v2 太阳与恒星天体物理
摘要
生成式用户界面（UI）为根据个体用户需求按需适应界面提供了新的机会，但由于可取的UI属性是主观的、难以表述，且从稀疏反馈中推断其成本较高，个性化仍然困难。我们通过一个包含20名训练设计师对同一600个生成UI进行成对判断的数据集来研究这一问题，使我们能够直接分析偏好差异。我们发现设计师之间存在显著的偏好不一致（平均kappa为0.25），且书面理由揭示了即使设计师诉诸类似的概念（如层次结构或干净），他们在如何定义、优先排序和应用这些概念方面存在差异。针对这些发现，我们开发了一种样本高效的个性化方法，该方法将新用户表示为相对于先前设计师，而不是基于固定的设计概念。 在技术评估中，我们的偏好模型均超过了预训练的UI评估器和更大的多模态模型，随着额外反馈的增加，其规模也更好。当用于个性化生成时，它还能产生被12名新设计师偏好于基线方法（包括直接用户提示）的界面。我们的发现表明，轻量级的偏好获取可作为面向个性化生成式UI系统的实用基础。
引用
@article{arxiv.2604.09875,
title = {Galactic Archaeology with the Subaru `\=Onohi`ula Prime Focus Spectrograph Strategic Program},
author = {Masashi Chiba and Rosemary F. G. Wyse and Evan N. Kirby and Judith G. Cohen and László Dobos and Roman Gerasimov and Miho N. Ishigaki and Kohei Hayashi and Carrie Filion and Magda Arnaboldi and Souradeep Bhattacharya and Yutaka Hirai and Chiaki Kobayashi and Yutaka Komiyama and Pete B. Kuzma and Itsuki Ogami and Ana L. Chies-Santos and Nicole L. Klock-Miranda and Federico Sestito and Tamás Budavári and Andrew P. Cooper and Keyi Ding and Ivanna Escala and Elisa G. M. Ferreira and Ortwin Gerhard and Lauren Henderson and Jihye Hong and Shunichi Horigome and Ryota Ikeda and Ryo Ishikawa and Takanobu Kirihara and Zhuohan Li and Nicolas Martin and Rin Miyazaki Sakurako Okamoto and Rohan Pattnaik and Kyosuke Sato and Yoshihisa Suzuki and Alexander S. Szalay and Dafa Wardana and Viska Wei and Wenbo Wu and Zhenyu Wu and Xinfeng Xu and Xianhao Ye and Yohei Miki and Xiangwei Zhang and Gang Zhao and Jingkun Zhao and Xiaosheng Zhao},
journal= {arXiv preprint arXiv:2604.09875},
year = {2026}
}
备注
The Galactic Archaeology science case for the Subaru Strategic Program for the `\=Onohi`ula Prime Focus Spectrograph. Not yet submitted to any journal. v2: Author list and affiliations updated