Junyi Zhu
← All publications

Adversarial Preference Learning for Robust LLM Alignment

Yuanfu Wang, Pengyu Wang, Chenyang Xi, Bo Tang, Junyi Zhu, Wenqiang Wei, Chen Chen, Chao Yang, Jingfeng Zhang, Chaochao Lu, Yijun Niu, Keming Mao, Zhiyu Li, Feiyu Xiong, Jie Hu, Mingchuan Yang

ACL Findings 2025Findings of the Association for Computational Linguistics: ACL 2025

Adversarial Preference Learning framework: an attacker model rewrites a harmful prompt into adversarial variants, a reward based on the defender's own preference probabilities scores each variant's attack effectiveness, the most and least effective rewrites train the attacker, and the most effective one drives the defender's preference update — a closed loop.

In brief

Adversarial Preference Learning (APL) hardens LLM alignment with an attacker that co-evolves with the model — no reward model, no human red-teaming needed. Three ideas drive it: the model's own preference probabilities serve as the harmfulness signal, sidestepping the reward-hacking risk of a separate reward model; a generative attacker rewrites each training prompt into its most damaging variants; and attacker and defender co-evolve in a closed loop, one discovering vulnerabilities as the other patches them. On Mistral-7B-Instruct, APL wins 83.33% of GPT-4o-judged harmlessness comparisons against the base model — versus 71.67% for standard preference tuning (DPO) — and cuts attack success by up to 65% on the HarmBench attack benchmark, with general utility close to baseline though over-refusals rise.

Key takeaways

Abstract

Modern language models often rely on Reinforcement Learning from Human Feedback (RLHF) to encourage safe behaviors. However, they remain vulnerable to adversarial attacks due to three key limitations: (1) the inefficiency and high cost of human annotation, (2) the vast diversity of potential adversarial attacks, and (3) the risk of feedback bias and reward hacking. To address these challenges, we introduce Adversarial Preference Learning (APL), an iterative adversarial training method incorporating three key innovations. First, a direct harmfulness metric based on the model’s intrinsic preference probabilities, eliminating reliance on external assessment. Second, a conditional generative attacker that synthesizes input-specific adversarial variations. Third, an iterative framework with automated closed-loop feedback, enabling continuous adaptation through vulnerability discovery and mitigation. Experiments on Mistral-7B-Instruct-v0.3 demonstrate that APL significantly enhances robustness, achieving 83.33% harmlessness win rate over the base model (evaluated by GPT-4o), reducing harmful outputs from 5.88% to 0.43% (measured by LLaMA-Guard), and lowering attack success rate by up to 65% according to HarmBench. Notably, APL maintains competitive utility, with an MT-Bench score of 6.59 (comparable to the baseline 6.78) and an LC-WinRate of 46.52% against the base model.

BibTeX

@inproceedings{apl,
  title     = {Adversarial Preference Learning for Robust {LLM} Alignment},
  author    = {Wang, Yuanfu and Wang, Pengyu and Xi, Chenyang and Tang, Bo and Zhu, Junyi and Wei, Wenqiang and Chen, Chen and Yang, Chao and Zhang, Jingfeng and Lu, Chaochao and Niu, Yijun and Mao, Keming and Li, Zhiyu and Xiong, Feiyu and Hu, Jie and Yang, Mingchuan},
  year      = {2025},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2025},
  pages     = {21865--21881},
  doi       = {10.18653/v1/2025.findings-acl.1126},
  url       = {https://doi.org/10.18653/v1/2025.findings-acl.1126},
}