Guardado en:
Detalles Bibliográficos
Autores principales: Hu, Zhanhao, Huang, Xiao, Mendoza, Patrick, Alghamdi, Emad A., Alomair, Basel, Popa, Raluca Ada, Wagner, David
Formato: Preprint
Publicado: 2026
Materias:
Acceso en línea:https://arxiv.org/abs/2605.14194
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866909041524998144
author Hu, Zhanhao
Huang, Xiao
Mendoza, Patrick
Alghamdi, Emad A.
Alomair, Basel
Popa, Raluca Ada
Wagner, David
author_facet Hu, Zhanhao
Huang, Xiao
Mendoza, Patrick
Alghamdi, Emad A.
Alomair, Basel
Popa, Raluca Ada
Wagner, David
contents Large Language Models (LLMs) pose a significant risk of safety misalignment after finetuning, as models can be compromised by both explicitly and implicitly harmful data. Even some seemingly benign data can inadvertently steer a model towards misaligned behaviors. To address this, we introduce GradShield, a principled filtering method that safeguards LLMs during finetuning by identifying and removing harmful data points before they corrupt the model's alignment. It removes potentially harmful data by computing a Finetuning Implicit Harmfulness Score (FIHS) for each data point and employs an adaptive thresholding algorithm. We apply GradShield to multiple utility fine-tuning tasks across varying levels of harmful data and evaluate the safety and utility performance of the resulting LLMs using various metrics. The results show that GradShield outperforms all baseline methods, consistently maintaining an Attack Success Rate (ASR) below $6\%$ while preserving utility performance.
format Preprint
id arxiv_https___arxiv_org_abs_2605_14194
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle GradShield: Alignment Preserving Finetuning
Hu, Zhanhao
Huang, Xiao
Mendoza, Patrick
Alghamdi, Emad A.
Alomair, Basel
Popa, Raluca Ada
Wagner, David
Computation and Language
Large Language Models (LLMs) pose a significant risk of safety misalignment after finetuning, as models can be compromised by both explicitly and implicitly harmful data. Even some seemingly benign data can inadvertently steer a model towards misaligned behaviors. To address this, we introduce GradShield, a principled filtering method that safeguards LLMs during finetuning by identifying and removing harmful data points before they corrupt the model's alignment. It removes potentially harmful data by computing a Finetuning Implicit Harmfulness Score (FIHS) for each data point and employs an adaptive thresholding algorithm. We apply GradShield to multiple utility fine-tuning tasks across varying levels of harmful data and evaluate the safety and utility performance of the resulting LLMs using various metrics. The results show that GradShield outperforms all baseline methods, consistently maintaining an Attack Success Rate (ASR) below $6\%$ while preserving utility performance.
title GradShield: Alignment Preserving Finetuning
topic Computation and Language
url https://arxiv.org/abs/2605.14194