Cover Image

Towards Understanding Safety Alignment: A Mechanistic Perspective from Safety Neurons

Fuente: arXiv

Saved in:

Bibliographic Details
Main Authors:	Chen, Jianhui, Wang, Xiaozhi, Yao, Zijun, Bai, Yushi, Hou, Lei, Li, Juanzi
Format:	Preprint
Published:	2024
Subjects:	Computation and Language Artificial Intelligence Machine Learning
Online Access:	Acceder al recurso
Tags:	Add Tag No Tags, Be the first to tag this record!

Similar Items

Guiding LLM Post-training Data Engineering with Model Internals from Sparse Autoencoders
by: Jing, Yi, et al.
Published: (2026)

NeuronTune: Fine-Grained Neuron Modulation for Balanced Safety-Utility Alignment in LLMs
by: Pan, Birong, et al.
Published: (2025)

WildReward: Learning Reward Models from In-the-Wild Human Interactions
by: Peng, Hao, et al.
Published: (2026)

A Single Neuron Is Sufficient to Bypass Safety Alignment in Large Language Models
by: Kazemi, Hamid, et al.
Published: (2026)

LongWriter-Zero: Mastering Ultra-Long Text Generation via Reinforcement Learning
by: Wu, Yuhao, et al.
Published: (2025)

Pre-training Distillation for Large Language Models: A Design Space Exploration
by: Peng, Hao, et al.
Published: (2024)

Multilingual Safety Alignment via Self-Distillation
by: Qin, Ruiyang, et al.
Published: (2026)

Boundary-Guided Policy Optimization for Memory-efficient RL of Diffusion Large Language Models
by: Lin, Nianyi, et al.
Published: (2025)

Test-Time Safety Alignment
by: Saglam, Baturay, et al.
Published: (2026)

Agentic Reward Modeling: Integrating Human Preferences with Verifiable Correctness Signals for Reliable Reward Systems
by: Peng, Hao, et al.
Published: (2025)

Safety Alignment Can Be Not Superficial With Explicit Safety Signals
by: Li, Jianwei, et al.
Published: (2025)

MESA: Improving MoE Safety Alignment via Decentralized Expertise
by: Sun, Yitong, et al.
Published: (2026)

Mechanistic Data Attribution: Tracing the Training Origins of Interpretable LLM Units
by: Chen, Jianhui, et al.
Published: (2026)

LongBench v2: Towards Deeper Understanding and Reasoning on Realistic Long-context Multitasks
by: Bai, Yushi, et al.
Published: (2024)

WaterBench: Towards Holistic Evaluation of Watermarks for Large Language Models
by: Tu, Shangqing, et al.
Published: (2023)

AdaptThink: Reasoning Models Can Learn When to Think
by: Zhang, Jiajie, et al.
Published: (2025)

Auxiliary Metrics Help Decoding Skill Neurons in the Wild
by: Zhao, Yixiu, et al.
Published: (2025)

Picky LLMs and Unreliable RMs: An Empirical Study on Safety Alignment after Instruction Tuning
by: Li, Guanlin, et al.
Published: (2025)

DeepPrune: Parallel Scaling without Inter-trace Redundancy
by: Tu, Shangqing, et al.
Published: (2025)

Lifelong Safety Alignment for Language Models
by: Wang, Haoyu, et al.
Published: (2025)

Course-Correction: Safety Alignment Using Synthetic Preferences
by: Xu, Rongwu, et al.
Published: (2024)

Superficial Safety Alignment Hypothesis
by: Li, Jianwei, et al.
Published: (2024)

Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications
by: Wei, Boyi, et al.
Published: (2024)

LASA: Language-Agnostic Semantic Alignment at the Semantic Bottleneck for LLM Safety
by: Yang, Junxiao, et al.
Published: (2026)

Unraveling and Mitigating Safety Alignment Degradation of Vision-Language Models
by: Liu, Qin, et al.
Published: (2024)

LongSafety: Enhance Safety for Long-Context LLMs
by: Huang, Mianqiu, et al.
Published: (2024)

Toward Understanding Unlearning Difficulty: A Mechanistic Perspective and Circuit-Guided Difficulty Metric
by: Cheng, Jiali, et al.
Published: (2026)

Safety Reasoning with Guidelines
by: Wang, Haoyu, et al.
Published: (2025)

Data Advisor: Dynamic Data Curation for Safety Alignment of Large Language Models
by: Wang, Fei, et al.
Published: (2024)

Interpretable Safety Alignment via SAE-Constructed Low-Rank Subspace Adaptation
by: Wang, Dianyun, et al.
Published: (2025)

Towards Unified Alignment Between Agents, Humans, and Environment
by: Yang, Zonghan, et al.
Published: (2024)

Model Merging and Safety Alignment: One Bad Model Spoils the Bunch
by: Hammoud, Hasan Abed Al Kader, et al.
Published: (2024)

CARE: Decoding Time Safety Alignment via Rollback and Introspection Intervention
by: Hu, Xiaomeng, et al.
Published: (2025)

Emulated Disalignment: Safety Alignment for Large Language Models May Backfire!
by: Zhou, Zhanhui, et al.
Published: (2024)

Disentangled Safety Adapters Enable Efficient Guardrails and Flexible Inference-Time Alignment
by: Krishna, Kundan, et al.
Published: (2025)

Align Once, Benefit Multilingually: Enforcing Multilingual Consistency for LLM Safety Alignment
by: Bu, Yuyan, et al.
Published: (2026)

Foundational Challenges in Assuring Alignment and Safety of Large Language Models
by: Anwar, Usman, et al.
Published: (2024)

SteeringSafety: A Systematic Safety Evaluation Framework of Representation Steering in LLMs
by: Siu, Vincent, et al.
Published: (2025)

Re-Emergent Misalignment: How Narrow Fine-Tuning Erodes Safety Alignment in LLMs
by: Giordani, Jeremiah
Published: (2025)

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection
by: Shairah, Harethah Abu, et al.
Published: (2025)