EMNLP Findings2025
Beyond Surface Alignment: Rebuilding LLMs Safety Mechanism via Probabilistically Ablating Refusal Direction
Yuanbo Xie, Yingjie Zhang, Tianyun Liu, Duohe Ma, Tingwen Liu
Abstract
Jailbreak attacks pose persistent threats to large language models (LLMs). Current safety alignment methods have attempted to address these issues, but they experience two significant limitations: insufficient safety alignment depth and unrobust internal defense mechanisms. These limitations make them vulnerable to adversarial attacks such as prefilling and refusal direction manipulation. We introduce DeepRefusal, a robust safety alignment framework that overcomes these issues. DeepRefusal forces the model to dynamically rebuild its refusal mechanisms from jailbreak states. This is achieved by probabilistically ablating the refusal direction across layers and token depths during fine-tuning. Our method not only defends against prefilling and refusal direction attacks but also demonstrates strong resilience against other unseen jailbreak strategies. Extensive evaluations on four open-source LLM families and six representative attacks show that DeepRefusal reduces attack success rates by approximately 95%, while maintaining model capabilities with minimal performance degradation.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
Lineage
Builds on20
- Measuring Massive Multitask Language UnderstandingDan Hendrycks, Collin Burns, Steven Basart, Andy Zou et al.ICLR 2021 · 7,905 citations
- Jailbroken: How Does LLM Safety Training Fail?Alexander Wei, Nika Haghtalab, Jacob SteinhardtNeurIPS 2023 · 2,230 citations
- Refusal in Language Models Is Mediated by a Single DirectionAndy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka et al.NeurIPS 2024 · 1,166 citations
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen et al.ICLR 2024 · 1,104 citations
- HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust RefusalMantas Mazeika, Long Phan, Xuwang Yin, Andy Zou et al.ICML 2024 · 1,031 citations
Related papers
- Improving LLM Safety Alignment with Dual-Objective OptimizationXuandong Zhao, Will Cai, Tianneng Shi, David Huang et al.ICML 2025
- Bleeding Pathways: Vanishing Discriminability in LLM Hidden States Fuels Jailbreak AttacksYingjie Zhang, Tong Liu, Zhe Zhao, Guozhu Meng et al.NDSS 2026 · 5 citations
- Safety Alignment Should be Made More Than Just a Few Tokens DeepXiangyu Qi, Ashwinee Panda, Kaifeng Lyu, Xiao Ma et al.ICLR 2025
- QueryAttack: Jailbreaking Aligned Large Language Models Using Structured Non-natural Query LanguageQingsong Zou, Jingyu Xiao, Qing Li, Zhi Yan et al.ACL Findings 2025 · 4 citations
- On the Vulnerability of Safety Alignment in Open-Access LLMsJingwei Yi, Rui Ye, Qisi Chen, Bin Zhu et al.ACL Findings 2024 · 15 citations
