Light Alignment Improves LLM Safety via Model Self-Reflection with a Single Neuron

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Shen, Sicheng, Lv, Mingyang, Shen, Han, Wu, Jialin, Wang, Binghao, Yang, Zhou, Shen, Guobin, Zhao, Dongcheng, Zhao, Feifei, Zeng, Yi
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911416397594624
author Shen, Sicheng
Lv, Mingyang
Shen, Han
Wu, Jialin
Wang, Binghao
Yang, Zhou
Shen, Guobin
Zhao, Dongcheng
Zhao, Feifei
Zeng, Yi
author_facet Shen, Sicheng
Lv, Mingyang
Shen, Han
Wu, Jialin
Wang, Binghao
Yang, Zhou
Shen, Guobin
Zhao, Dongcheng
Zhao, Feifei
Zeng, Yi
contents The safety of large language models (LLMs) has increasingly emerged as a fundamental aspect of their development. Existing safety alignment for LLMs is predominantly achieved through post-training methods, which are computationally expensive and often fail to generalize well across different models. A small number of lightweight alignment approaches either rely heavily on prior-computed safety injections or depend excessively on the model's own capabilities, resulting in limited generalization and degraded efficiency and usability during generation. In this work, we propose a safety-aware decoding method that requires only low-cost training of an expert model and employs a single neuron as a gating mechanism. By effectively balancing the model's intrinsic capabilities with external guidance, our approach simultaneously preserves utility and enhances output safety. It demonstrates clear advantages in training overhead and generalization across model scales, offering a new perspective on lightweight alignment for the safe and practical deployment of large language models. Code: https://github.com/Beijing-AISI/NGSD.
format Preprint
id arxiv_https___arxiv_org_abs_2602_02027
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Light Alignment Improves LLM Safety via Model Self-Reflection with a Single Neuron
Shen, Sicheng
Lv, Mingyang
Shen, Han
Wu, Jialin
Wang, Binghao
Yang, Zhou
Shen, Guobin
Zhao, Dongcheng
Zhao, Feifei
Zeng, Yi
Artificial Intelligence
Machine Learning
The safety of large language models (LLMs) has increasingly emerged as a fundamental aspect of their development. Existing safety alignment for LLMs is predominantly achieved through post-training methods, which are computationally expensive and often fail to generalize well across different models. A small number of lightweight alignment approaches either rely heavily on prior-computed safety injections or depend excessively on the model's own capabilities, resulting in limited generalization and degraded efficiency and usability during generation. In this work, we propose a safety-aware decoding method that requires only low-cost training of an expert model and employs a single neuron as a gating mechanism. By effectively balancing the model's intrinsic capabilities with external guidance, our approach simultaneously preserves utility and enhances output safety. It demonstrates clear advantages in training overhead and generalization across model scales, offering a new perspective on lightweight alignment for the safe and practical deployment of large language models. Code: https://github.com/Beijing-AISI/NGSD.
title Light Alignment Improves LLM Safety via Model Self-Reflection with a Single Neuron
topic Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2602.02027