method · first seen 4 Oct 2026
safety alignment
Also: refusal alignment, LLM safety alignment
Training language models to refuse harmful inputs and behave according to safety constraints.
Safety alignment means training language models to refuse harmful requests and follow safety constraints, typically through supervised fine-tuning, RLHF or preference optimization. The sources are a steady stream of academic papers on familiar sub-problems: shallow or sparse alignment that breaks under pruning, fine-tuning or context compression; safety gaps in low-resource languages such as Hausa; interpretability probes of what stays hidden behind refusals; and alignment for reasoning models using verifiable rewards. One repository releases code for a NeurIPS paper on category-adaptive DPO. The volume reflects an active, mature research area rather than a new idea.
The concept has been central to LLM development since RLHF-era models. The current work matters because it documents how fragile refusal behavior is under fine-tuning, translation and compression, which affects anyone shipping open-weight or adapted models.
Analyst note (established), 18 Nov 2026.
Mentions per week
PapersCodeDiscussion
Rank 101 on 15 Nov 2026 · 12 mentions in 28 days · 12 independent groups · papers → code · first seen 6 weeks ago · acceleration +0.10
Earliest sources
- Beyond Refusal Patterns: Safe-Role Internalization for Robust and Generalizable LLM Safety AlignmentPapers · 4 Oct 2026 · Jinghao Pang, Jitai Hao, Qiang Huang, Zhaochun Ren, Jun Yu
- SafeEvo: Deciphering the Safety Alignment Mechanism and Evolution in Language ModelsPapers · 7 Oct 2026 · Miao Yu, Hao Huang, Lu Yuan, Yunpeng Li, Kun Wang, Zuming Jiang
- DANGERLENS: Probing Dangerous Representations Hidden by Safety AlignmentPapers · 14 Oct 2026 · Tianzhuo Yang, Shaoheng Yan, Lin Lu, Jiawei Chen, Kaile Wang, Fanyu Meng, Minglu…
- In the Guise of Correction: How Framing and Cultural Adaptation Erode LLM Safety on Real-World Criminal CasesPapers · 15 Oct 2026 · Mohsinul Kabir, Md. Tariquzzaman, Karam S. R. Elradie, Polydoros Giannouris, Yue…
- Rethinking GRPO for Large Reasoning Model Safety AlignmentPapers · 16 Oct 2026 · Jiawei Chen, Yi Zhang, Jiawei Du, Zhonghou Lv, Tairan Chen, Junyuan Shang, Jinch…
- When Safety Fails Collectively: Robust Alignment Against Harmful Fine-Tuning via Joint Weight PerturbationsPapers · 19 Oct 2026 · Zhiyuan Wang, Yuezhi Zhou, Zhu Zhang, Zhixing Tan
Latest sources
- Data-Efficient LLM Safety Alignment Guided by Interpretable FeaturesPapers · 12 Nov 2026 · Hongyu Li, Zhan Yu, Zhe Zhao, Kimmo Yan, Xiong Kuang, Renfen Hu
- RaymondY/Cat-DPOCode · 7 Nov 2026 · RaymondY · 75 stars
- When Safety Becomes An Outlier: Understanding the Retention of LLM Safety BehaviorsPapers · 7 Nov 2026 · Binchi Zhang, Hadi Abdullah, Yiwei Cai, Jundong Li
- Safety Direction Adapter: A Training-Free Post-Hoc Safety Patch for Fine-Tuned Open-Weight LLMsPapers · 6 Nov 2026 · Chansu Han, Takeshi Takahashi
- SAVeR$^2$: Reasoning-based Safety Alignment for Large Reasoning Models via Verifiable RewardsPapers · 5 Nov 2026 · Yichen Sun, Qirui Liu, Yan Wang, Jianan Lin, Linbo Jiang, Shiyu Wang, Zhibo Wang…
- Addition by Subtraction: Toward Mitigating Fragile Safety Alignment in Large Language ModelsPapers · 4 Nov 2026 · Hamid Dashtbani, Alexandra Dmitrienko
- Does Context Compression Preserve Refusal Alignment?Papers · 4 Nov 2026 · Anthony Nguyen, Wenjun Lin
- Safin-1.5: Safety-Native Architecture and Verifiable Reasoning for Recursive Safety ImprovementPapers · 29 Oct 2026 · Jingnan Zheng, Cheng Jin, Ming Zhang, Huan Sun, Zhongtian Ma, Kaisen Yang, Ziyua…
- Aligning Large Language Models for Safety in African Languages: A Case Study of HausaPapers · 27 Oct 2026 · Valdini Douglace Lemofouet, Blessing Ngozi Uzor, Paula Chikaodinaka Anyanwu, Dan…
- Global Security Blind Spots: Translation as a Safety Bypass in Low-Resource LanguagesPapers · 19 Oct 2026 · Jalil Alva, David Harper, Abigail E. Thornton, Rafael S. Villalobos Alva, Leonar…
- Safety Mid-Training: Internalizing LLM Safety as a Foundational CapabilityPapers · 19 Oct 2026 · Zhengyue Zhao, Liwei Jiang, Yejin Choi, Chaowei Xiao
- When Safety Fails Collectively: Robust Alignment Against Harmful Fine-Tuning via Joint Weight PerturbationsPapers · 19 Oct 2026 · Zhiyuan Wang, Yuezhi Zhou, Zhu Zhang, Zhixing Tan