LLM Safety Alignment

LLM: Large Language Model

Large language model safety alignment studies how to make models follow intended rules, avoid enabling harm, and remain useful when requests are ambiguous or adversarial. Common approaches include supervised safety fine-tuning and preference optimization such as reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO), adversarial red-teaming, and inference-time controls such as refusal steering, prompt routing, or guard models. Evaluations use jailbreak and harmful-request suites such as HarmBench and JailbreakBench, alongside tests of over-refusal, multilingual and multimodal inputs, fine-tuning drift, and tool-using agents. A central challenge is achieving robust, context-sensitive safety without blocking benign help: refusals can fail under new formats, languages, or deployment pressures, while stronger safeguards can reduce utility and do not always predict real agent behavior.

Latest papers

347 papers40 papers in the last four weeks, up 38% on the four weeks before. 0.2% of all new papers.

ShallowDefend: Harnessing Shallow Safety Alignment for Multi-Turn Defender
Oct 29, 2026Qiyao Ma, Xiaodong Liu, Weiwei Yang +9LLM Safety Alignment Jailbreak Defense
University of California, Davis · Microsoft Research
How I Aligned Your Model: Scaling Up Multilingual Safety Training with Templates
Oct 29, 2026Max Neuwinger, Hanna Yukhymenko, Imanol SchlagLLM Safety Alignment Language Model Safety Evaluation
ETH Zurich
Teach reality, not ethics: a drop-in continued-pretraining corpus installs coupled human-biosphere-AI welfare dispositions across model families
Oct 27, 2026Trent R. Northen, Mingxun WangLLM Safety Alignment
Bioaligned Labs
Aligning Large Language Models for Safety in African Languages: A Case Study of Hausa
Oct 27, 2026Valdini Douglace Lemofouet, Blessing Ngozi Uzor, Paula Chikaodinaka Anyanwu +4LLM Safety Alignment Language Model Safety Evaluation
African Institute for Mathematical Sciences (AIMS), Cameroon · Bayero University Kano · University of Hamburg +1
MedConstBench: Benchmarking How Constitutions Steer LLMs in Medical Ethics
Oct 27, 2026Yibo Yang, Zerui Huang, Haomiao Yang +1Medical LLMs Moral Reasoning in Language Models
Independent Researcher, China · Xinjin District People’s Hospital, China · University of Exeter, United Kingdom
Improving Instruction Hierarchy by Learning from Goal-Goal Conflicts
Oct 26, 2026Sizhe Chen, Sanjay Kariyappa, Liwei Jiang +4LLM Safety Alignment Instruction Tuning
NVIDIA · UC Berkeley
When Personas Become Harmful: Guiding LLMs with Global Citizenship Values Alignment
Oct 24, 2026Siqi Shen, Lajanugen Logeswaran, Moontae Lee +2LLM Safety LLM Safety Alignment
University of Michigan · LG AI Research · University of Illinois at Chicago
Auditing Safety Preference Supervision with Recorded Derivations
Oct 19, 2026Chenhan Lyu, Amir RahmaniLLM Auditing LLM Safety Alignment
University of California, Irvine Irvine, USA
Safety Mid-Training: Internalizing LLM Safety as a Foundational Capability
Oct 19, 2026Zhengyue Zhao, Liwei Jiang, Yejin Choi +1Safety-Utility Trade-Off in Language Models LLM Safety Alignment
Johns Hopkins University · University of Washington · Stanford University
Training Natural Schemers: RL Can Amplify Oversight-Conditional Behavior
Oct 19, 2026Nathaniel Mitrani Hadida, Bart Jaworski, Fred Bruford +1Strategic Reasoning Risks in Language Models LLM Safety Alignment
MATS Research · Google DeepMind
Rethinking LLM Fine-Tuning via Weight Space Reparameterization: Preserving Safety during Downstream Adaptation
Oct 19, 2026Min-Seong Kim, Jongbok Won, Jeesup Park +2Safety-Preserving Fine-Tuning of Language Models LLM Fine-Tuning
Yonsei University
Safety Offset Learning: When Benign Fine-Tuning Misaligns LLMs
Oct 19, 2026Jesson Wang, Zhanhao Hu, Minghao Lin +3LLM Fine-Tuning LLM Safety Alignment
University of Southern California · University of California, Berkeley · University of Washington
Rethinking GRPO for Large Reasoning Model Safety Alignment
Oct 16, 2026Jiawei Chen, Yi Zhang, Jiawei Du +9Safety-Preserving Fine-Tuning of Language Models LLM Safety Alignment
East China Normal University · Baidu Inc. · Zhongguancun Academy +5
Identify, Classify, Act: Per-Instruction Controlled Compliance for LLMs over External Data
Oct 16, 2026Zesheng Shi, Xue Li, Chenrui Dai +1Prompt Injection Defense LLM Safety Alignment
Harbin Institute of Technology, Shenzhen, China
ContextLift: Breaking the Spurious Correlation Behind Conditional Emergent Misalignment
Oct 15, 2026Florian Mai, David KaczérEmergent Misalignment in Language Models LLM Safety Alignment
Bonn-Aachen International Center for Information Technology, University of Bonn, Germany · Lamarr Institute for Machine Learning and Artificial Intelligence, Germany
Resisting Harmful Instructions through Training-Free Weight Editing
Oct 15, 2026Ali Ebrahimpour-Boroojeny, Xiaocong Yang, Samuel Schapiro +4LLM Safety LLM Safety Alignment
Manacá-1B-Instruct-2: Reproducible Safety Alignment for a Native Brazilian-Portuguese LLM. Negative DPO Results, a Controlled Composition Scan, and a Silent Failure Mode of the LLM Judge
Oct 14, 2026Bruno Leonardo Santos Menezes, Carlos Leonardo Souza Cardoso, Fabio André Machado PortoPortuguese Language Models LLM Safety Alignment
National Laboratory for Scientific Computing (LNCC), Institute of Artificial Intelligence, Petrópolis, Brazil
DANGERLENS: Probing Dangerous Representations Hidden by Safety Alignment
Oct 14, 2026Tianzhuo Yang, Shaoheng Yan, Lin Lu +5LLM Safety LLM Safety Alignment
Institute for Artificial Intelligence, Peking University · JIUTIAN Research, China Mobile
Scenes as an Attack Surface: Uncovering Scene-Grounded Safety Failures in MLLMs
Oct 8, 2026Tongxi Wu, Yuyang Cheng, Jiacheng Xian +5MLLM Safety Adversarial Attacks on VLMs
Nanjing University · The Hong Kong University of Science and Technology (Guangzhou) · Independent Researcher +2
ReSI: Recursive Safety Improvement toward Resistant and Resilient AI
Oct 8, 2026Jingnan Zheng, Dongcheng Zhang, Yi Zhang +10LLM Safety Alignment Recursive Self-Improvement
Shanghai AI Laboratory · National University of Singapore · Shanghai Jiao Tong University +2
Distillation for Incrimination and Distillation for Capabilities
Oct 7, 2026Sebastian Prasanna, Jacqueline Tay, Alek WestoverLLM Safety Alignment Language Model Distillation
Redwood Research
How Narrative Wrapping Affects LLM Refusal: A Cross-Language Benchmark and Defense
Oct 7, 2026Zhankai Ye, Yanning Wang, Yukai Jin +5Refusal Behavior in Language Models LLM Safety Alignment
Florida State University · Texas Christian University · University of Pennsylvania +1
SLDR: Defending Against Malicious Fine-tuning via Selective Layers Recovery and Dynamic Routing
Oct 7, 2026Hui Zhang, Yachao Yuan, Jiayun Wang +3LLM Security LLM Safety
North China Electric Power University · Southeast University · Soochow University +1
Safe at One Loop, Risky at Another: Aligning Safety Across Recurrent Depths in Looped Language Models
Oct 7, 2026Yi Wang, Xiuyuan Qi, Dongqi Han +2LLM Safety Alignment Language Model Safety Evaluation
ShanghaiTech University · Microsoft Research Asia
SafeEvo: Deciphering the Safety Alignment Mechanism and Evolution in Language Models
Oct 7, 2026Miao Yu, Hao Huang, Lu Yuan +3Refusal Mechanisms in Language Models LLM Safety Alignment
The University of Hong Kong (HKU) · Chinese Academy of Sciences (CAS) · Information Engineering University (IEU) +1
SIGMA: Self-Improving Alignment Generalization from a Model Spec
Oct 6, 2026Jingyu Zhang, Shruti Palaskar, Daniel Khashabi +3LLM Safety Alignment LLM Agent Safety
Apple · Johns Hopkins University
The Model Plants the Trigger: Answer-Side Backdoor Attacks in Multi-Turn Large Language Models
Oct 6, 2026Yibo Zhang, Tianrong Guan, Liang Lin +3LLM Backdoor Attacks Backdoor Attacks
Queen Mary University of London, UK · Squirrel AI Learning, USA
Beyond Refusal Patterns: Safe-Role Internalization for Robust and Generalizable LLM Safety Alignment
Oct 4, 2026Jinghao Pang, Jitai Hao, Qiang Huang +2LLM Alignment LLM Safety Alignment
Harbin Institute of Technology (Shenzhen) · Leiden University
Safe Context Switching for Agents in the Wild: Mitigating Subspace Interference via Orthogonal Adaptation
Oct 4, 2026Akash Das, Ishan RoyLLM Safety Alignment LLM Safety
Fidelity Investments
Reactivating Alignment: Defending LLMs from Jailbreaks via Intention-Aware Input-Output Matching
Oct 3, 2026Luoyu Chen, Weiqi Wang, Chenhan Zhang +2Jailbreak Defense LLM Safety Alignment
University of Technology Sydney, Sydney, NSW, Australia · Xi’an Jiaotong University, Xi’an, Shaanxi, China · Southeast University, Nanjing, Jiangsu, China

Oct 29, 2026J/K move · Enter open · S save

ShallowDefend: Harnessing Shallow Safety Alignment for Multi-Turn Defender

Qiyao Ma, Xiaodong Liu, Weiwei Yang, Mingqian Feng+8

University of California, Davis · Microsoft Research

LLM Safety Alignment Jailbreak Defense Safety-Preserving Fine-Tuning of Language Models

Open paper AskPDF arXiv

Safety alignment has enabled large language models to reliably refuse harmful requests directly, yet the same request distributed across a multi-turn conversation still succeeds far too often. This brittleness has a mechanistic root: Alignment is shallow, with safety behavior concentrated in a model's first few output tokens, and existing defenses inherit the weakness, supervising the surface form of the response without ever giving the model an explicit safety decision to learn. We take the opposite view of shallowness: if a short span determines the trajectory of an entire response, that span is the highest-leverage place to intervene, and the defender can occupy it first. We propose Assess

\to

Decide

\to

Respond, a framework in which the defender assesses the user's intent, commits to an explicit, parseable safety decision, and only then generates a reply conditioned on it, with the commitment trained by reinforcement learning at the payload turn. Because the decision is a token span, the reward can be a rule: a self-verified reward checks the committed label against the conversation's known nature, requiring no judge and no annotation, while the same format also supports an LLM-as-judge reward over a graded decision space with safe redirection. Across two model families and fourteen benchmarks spanning safety, utility, and helpfulness, our defenders achieve the strongest multi-turn safety in a controlled comparison, maintaining their lead in safety while preserving utility at a modest cost to helpfulness. A per-token KL analysis and interventional probes confirm why this works: training concentrates its effect on the brief deliberation that sets a response's trajectory, and the decision committed there is a causal control point. The very shallowness attackers exploit becomes the mechanism of defense.