CARE: Decoding-Time Safety Alignment via Rollback and Introspection Intervention
Xiaomeng Hu, Fei Huang, Chenhan Yuan, Junyang Lin, Tsung-Yi Ho
Abstract
As large language models (LLMs) are increasingly deployed in real-world applications, ensuring the safety of their outputs during decoding has become a critical challenge. However, existing decoding-time interventions, such as Contrastive Decoding, often force a severe trade-off between safety and response quality. In this work, we propose CARE, a novel framework for decoding-time safety alignment that integrates three key components: (1) a guard model for real-time safety monitoring, enabling detection of potentially unsafe content; (2) a rollback mechanism with a token buffer to correct unsafe outputs efficiently at an earlier stage without disrupting the user experience; and (3) a novel introspection-based intervention strategy, where the model generates self-reflective critiques of its previous outputs and incorporates these reflections into the context to guide subsequent decoding steps. The framework achieves a superior safety-quality trade-off by using its guard model for precise interventions, its rollback mechanism for timely corrections, and our novel introspection method for effective self-correction. Experimental results demonstrate that our framework achieves a superior balance of safety, quality, and efficiency, attaining a low harmful response rate and minimal disruption to the user experience while maintaining high response quality.
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.
1 top-tier paper has cited it since. Know what they found before your reviewers do.
Lune can draw on
the full paper · the top-tier papers that cite it · related work
Free to start. No credit card required.
Lineage
- Tree of Attacks: Jailbreaking Black-Box LLMs Automatically cites AutoDAN: Generating Stealthy Jailbreak Prompts on Aligned Large Language Models
- AutoDAN: Generating Stealthy Jailbreak Prompts on Aligned Large Language Models cites Training language models to follow instructions with human feedback
- ARGS: Alignment as Reward-Guided Search cites Training language models to follow instructions with human feedback
Cited by top-tier papers1
Builds on10
- Training language models to follow instructions with human feedbackLong Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida et al.NeurIPS 2022 · 24,707 citations
- Tree of Attacks: Jailbreaking Black-Box LLMs AutomaticallyAnay Mehrotra, Manolis Zampetakis, Paul Kassianik, Blaine Nelson et al.NeurIPS 2024 · 835 citations
- AutoDAN: Generating Stealthy Jailbreak Prompts on Aligned Large Language ModelsXiaogeng Liu, Nan Xu, Muhao Chen, Chaowei XiaoICLR 2024 · 722 citations
- Many-shot JailbreakingCem Anil, Esin Durmus, Nina Panickssery, Mrinank Sharma et al.NeurIPS 2024 · 338 citations
- ARGS: Alignment as Reward-Guided SearchMaxim Khanov, Jirayu Burapacheep, Yixuan LiICLR 2024 · 101 citations
Related papers
- Safety Alignment of Large Language Models via Contrasting Safe and Harmful DistributionsXiaoyun Zhang, Zhengyue Zhao, Wenxuan Shi, Kaidi Xu et al.AAAI 2026 · 4 citations
- Root Defense Strategies: Ensuring Safety of LLM at the Decoding LevelXinyi Zeng, Yuying Shang, Jiawei Chen, Jingyuan Zhang et al.ACL 2025 · 6 citations
- SecDecoding: Steerable Decoding for Safer LLM GenerationJiayou Wang, Rundong Liu, Yue Hu, Huijia Wu et al.EMNLP Findings 2025
- SafeSwitch: Steering Unsafe LLM Behavior via Internal Activation SignalsPeixuan Han, Cheng Qian, Xiusi Chen, Yuji Zhang et al.EMNLP Findings 2025 · 3 citations
- ROSE Doesn't Do That: Boosting the Safety of Instruction-Tuned Large Language Models with Reverse Prompt Contrastive DecodingQihuang Zhong, Liang Ding, Juhua Liu, Bo Du et al.ACL Findings 2024 · 9 citations
