Towards Identification and Intervention of Safety-Critical Parameters in Large Language Models
2026-04-09 • Cryptography and Security
Cryptography and Security
AI summaryⓘ
The authors propose a new way called Expected Safety Impact (ESI) to understand which parts of large language models (LLMs) affect their safety the most. They find that important safety-related parameters differ between dense LLMs and Mixture-of-Experts (MoE) models, located in different model layers. Using ESI, they develop two methods: one to improve unsafe models by updating a few key parameters, and another to protect safe models when adding new skills. Their tests show these methods can effectively make models safer or keep them safe without hurting performance.
Large Language Models (LLMs)Expected Safety Impact (ESI)Dense ModelsMixture-of-Experts (MoE)Safety Enhancement Tuning (SET)Safety Preserving Adaptation (SPA)Value MatricesMLP LayersInstruction Fine-TuningAttack Success Rate
Authors
Weiwei Qi, Zefeng Wu, Tianhang Zheng, Zikang Zhang, Xiaojun Jia, Zhan Qin, Kui Ren
Abstract
Ensuring Large Language Model (LLM) safety is crucial, yet the lack of a clear understanding about safety mechanisms hinders the development of precise and reliable methodologies for safety intervention across diverse tasks. To better understand and control LLM safety, we propose the Expected Safety Impact (ESI) framework for quantifying how different parameters affect LLM safety. Based on ESI, we reveal distinct safety-critical patterns across different LLM architectures: In dense LLMs, many safety-critical parameters are located in value matrices (V) and MLPs in middle layers, whereas in Mixture-of-Experts (MoE) models, they shift to the late-layer MLPs. Leveraging ESI, we further introduce two targeted intervention paradigms for safety enhancement and preservation, i.e., Safety Enhancement Tuning (SET) and Safety Preserving Adaptation (SPA). SET can align unsafe LLMs by updating only a few safety-critical parameters, effectively enhancing safety while preserving original performance. SPA safeguards well-aligned LLMs during capability-oriented intervention (e.g., instruction tuning) by preventing disruption of safety-critical weights, allowing the LLM to acquire new abilities and maintain safety capabilities. Extensive evaluations on different LLMs demonstrate that SET can reduce the attack success rates of unaligned LLMs by over 50% with only a 100-iteration update on 1% of model weights. SPA can limit the safety degradation of aligned LLMs within 1% after a 1,000-iteration instruction fine-tuning on different tasks. Our code is available at: https://github.com/ZJU-LLM-Safety/SafeWeights-ACL.