@inproceedings{qiyuan-etal-2025-efficient,
    title = "Efficient Safety Alignment of Large Language Models via Preference Re-ranking and Representation-based Reward Modeling",
    author = "Qiyuan, Deng  and
      Bai, Xuefeng  and
      Chen, Kehai  and
      Wang, Yaowei  and
      Nie, Liqiang  and
      Zhang, Min",
    editor = "Che, Wanxiang  and
      Nabende, Joyce  and
      Shutova, Ekaterina  and
      Pilehvar, Mohammad Taher",
    booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2025",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingestion-acl-25/2025.acl-long.1504/",
    pages = "31156--31171",
    ISBN = "979-8-89176-251-0",
    abstract = "Reinforcement Learning (RL) algorithms for safety alignment of Large Language Models (LLMs), such as Direct Preference Optimization (DPO), encounter the challenge of distribution shift. Current approaches typically address this issue through online sampling from the target policy, which requires significant computational resources.In this paper, we hypothesize that during off-policy training, while the ranking order of output generated by policy changes, their overall distribution remains relatively stable.This stability allows the conversion of the sampling process from the target policy into a computationallyefficient re-ranking of preference data.Building on this hypothesis, we propose a new framework that leverages the model{'}s intrinsic safety judgment capability to extract reward signals, which are then used to calculate label confidence for preference reordering. Extensive experiments and theoretical analysis demonstrate that the proposed method effectively addresses the distribution shift issue, remarkably enhancing the safety performance while avoiding about 300x computational overheads."
}