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

2026-02-02

AI Summary

Current safety alignment methods for large language models either incur high computational costs and poor generalization or rely on predefined rules and the model’s intrinsic capabilities, struggling to balance efficiency and universality. This work proposes a safety-aware decoding mechanism based on single-neuron gating, which dynamically integrates the model’s internal reasoning with external safety guidance during generation through lightweight expert model training, enabling efficient self-reflective safety control. Remarkably, the approach requires only a single neuron to make safety decisions and consistently outperforms existing lightweight alignment strategies across multiple model scales. It significantly enhances safety while preserving output utility, substantially reduces training overhead, and demonstrates strong cross-model generalization capability.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Learning to Search

Application Category

• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

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.

Problem

• LLM safety
• alignment
• lightweight methods
• generalization
• computational cost

Innovation

• lightweight alignment
• single neuron gating
• safety-aware decoding
• model self-reflection
• LLM safety