EMNLP Findings2025
SafeSwitch: Steering Unsafe LLM Behavior via Internal Activation Signals
Peixuan Han, Cheng Qian, Xiusi Chen, Yuji Zhang, Heng Ji, Denghui Zhang
Abstract
Large language models (LLMs) exhibit exceptional capabilities across various tasks but also pose risks by generating harmful content.Existing safety mechanisms, while improving model safety, often lead to overly cautious behavior and fail to fully leverage LLMs' internal cognitive processes.Inspired by humans' reflective thinking capability, we first show that LLMs can similarly perform internal assessments about safety in their internal states.Building on this insight, we propose SafeSwitch, a dynamic framework that regulates unsafe outputs by utilizing the prober-based internal state monitor that actively detects harmful intentions, and activates a safety head that leads to safer and more conservative responses only when necessary.SafeSwitch reduces harmful outputs by approximately 80% on harmful queries while maintaining strong utility, reaching a Pareto optimal among several methods.Our method is also advantageous over traditional methods in offering more informative, context-aware refusals, and achieves these benefits while only tuning less than 6% of the original parameters.SafeSwitch demonstrates large language models' capacity for self-awareness and reflection regarding safety, offering a promising approach to more nuanced and effective safety controls.Codes for this work are available at https:// github.com/Hanpx20/SafeSwitch.
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.
6 top-tier papers have 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
- Rescorla-Wagner Steering of LLMs for Undesired Behaviors over Disproportionate Inappropriate Context cites Locating and Editing Factual Associations in GPT
- LLMs Encode Harmfulness and Refusal Separately cites Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!
- LLMs Encode Harmfulness and Refusal Separately cites Refusal in Language Models Is Mediated by a Single Direction
- Language Models Are Capable of Metacognitive Monitoring and Control of Their Internal Activations cites Refusal in Language Models Is Mediated by a Single Direction
Cited by top-tier papers6
- LLMs Encode Harmfulness and Refusal SeparatelyJiachen Zhao, Jing Huang, Zhengxuan Wu, David Bau et al.NeurIPS 2025 · 93 citations
- Language Models Are Capable of Metacognitive Monitoring and Control of Their Internal ActivationsJi-An Li, Huadong Xiong, Robert C. Wilson, Marcelo G. Mattar et al.NeurIPS 2025 · 49 citations
- Self-Aligned Reward: Towards Effective and Efficient ReasonersPeixuan Han, ADIT KRISHNAN, Gerald Friedland, Jiaxuan You et al.ICLR 2026 · 10 citations
- GAVEL: Towards Rule-Based Safety through Activation MonitoringShir Rozenfeld, Rahul Pankajakshan, Itay Zloczower, Eyal Lenga et al.ICLR 2026 · 4 citations
- Rescorla-Wagner Steering of LLMs for Undesired Behaviors over Disproportionate Inappropriate ContextRushi Wang, Jiateng Liu, Cheng Qian, Yifan Shen et al.EMNLP 2025
Builds on24
- Tree of Thoughts: Deliberate Problem Solving with Large Language ModelsShunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran et al.NeurIPS 2023 · 5,068 citations
- Locating and Editing Factual Associations in GPTKevin Meng, David Bau, Alex Andonian, Yonatan BelinkovNeurIPS 2022 · 3,415 citations
- Let's Verify Step by StepHunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards et al.ICLR 2024 · 3,045 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
Related papers
- ProSafePrune: Projected Safety Pruning for Mitigating Over-Refusal in LLMsZijun Chen, Wenbo Hu, Ya Li, Lei Miao et al.ICLR 2026
- CARE: Decoding-Time Safety Alignment via Rollback and Introspection InterventionXiaomeng Hu, Fei Huang, Chenhan Yuan, Junyang Lin et al.NeurIPS 2025 · 5 citations
- Dynamic Evaluation with Cognitive Reasoning for Multi-turn Safety of Large Language ModelsLanxue Zhang, Yanan Cao, Yuqiang Xie, Fang Fang et al.ACL 2025
- AlphaAlign: Incentivizing Safety Alignment with Extremely Simplified Reinforcement LearningYi Zhang, An Zhang, XiuYu Zhang, Leheng Sheng et al.ICLR 2026 · 15 citations
- Think Twice, Generate Once: Safeguarding by Progressive Self-ReflectionHoang Phan, Victor Li, Qi LeiEMNLP Findings 2025 · 1 citation
