Interpretable Safety Alignment via SAE-Constructed Low-Rank Subspace Adaptation

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Wang, Dianyun, Ma, Qingsen, Shang, Yuhu, Lu, Zhifeng, Xu, Zhenbo, Ning, Lechen, Wu, Huijia, He, Zhaofeng
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866911354751811584
author Wang, Dianyun
Ma, Qingsen
Shang, Yuhu
Lu, Zhifeng
Xu, Zhenbo
Ning, Lechen
Wu, Huijia
He, Zhaofeng
author_facet Wang, Dianyun
Ma, Qingsen
Shang, Yuhu
Lu, Zhifeng
Xu, Zhenbo
Ning, Lechen
Wu, Huijia
He, Zhaofeng
contents Safety alignment -- training large language models (LLMs) to refuse harmful requests while remaining helpful -- is critical for responsible deployment. Prior work established that safety behaviors are governed by low-rank structures, suggesting parameter-efficient fine-tuning (PEFT) should be well-suited for alignment. However, Low-Rank Adaptation (LoRA) consistently underperforms full fine-tuning and reinforcement learning on safety benchmarks. We attribute this gap to semantic entanglement: safety-relevant directions are intertwined with unrelated concepts due to polysemanticity, impeding implicit subspace identification. To address this, we propose SAILS (Safety Alignment via Interpretable Low-rank Subspace), which leverages Sparse Autoencoders (SAEs) to disentangle representations into monosemantic features, constructs an interpretable safety subspace from SAE decoder directions, and uses it to initialize LoRA adapters. Theoretically, we prove that SAE-based identification achieves arbitrarily small recovery error under monosemanticity assumptions, while direct identification suffers an irreducible error floor. Empirically, SAILS achieves up to 99.6% safety rate on Gemma-2-9B -- exceeding full fine-tuning by 7.4 points and matching RLHF-based models -- while updating only 0.19% of parameters and providing interpretability.
format Preprint
id arxiv_https___arxiv_org_abs_2512_23260
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Interpretable Safety Alignment via SAE-Constructed Low-Rank Subspace Adaptation
Wang, Dianyun
Ma, Qingsen
Shang, Yuhu
Lu, Zhifeng
Xu, Zhenbo
Ning, Lechen
Wu, Huijia
He, Zhaofeng
Computation and Language
Artificial Intelligence
Machine Learning
Safety alignment -- training large language models (LLMs) to refuse harmful requests while remaining helpful -- is critical for responsible deployment. Prior work established that safety behaviors are governed by low-rank structures, suggesting parameter-efficient fine-tuning (PEFT) should be well-suited for alignment. However, Low-Rank Adaptation (LoRA) consistently underperforms full fine-tuning and reinforcement learning on safety benchmarks. We attribute this gap to semantic entanglement: safety-relevant directions are intertwined with unrelated concepts due to polysemanticity, impeding implicit subspace identification. To address this, we propose SAILS (Safety Alignment via Interpretable Low-rank Subspace), which leverages Sparse Autoencoders (SAEs) to disentangle representations into monosemantic features, constructs an interpretable safety subspace from SAE decoder directions, and uses it to initialize LoRA adapters. Theoretically, we prove that SAE-based identification achieves arbitrarily small recovery error under monosemanticity assumptions, while direct identification suffers an irreducible error floor. Empirically, SAILS achieves up to 99.6% safety rate on Gemma-2-9B -- exceeding full fine-tuning by 7.4 points and matching RLHF-based models -- while updating only 0.19% of parameters and providing interpretability.
title Interpretable Safety Alignment via SAE-Constructed Low-Rank Subspace Adaptation
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2512.23260