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

TL;DR

This paper introduces SAILS, a method that uses sparse autoencoders to disentangle safety-related features in language models, enabling efficient and interpretable safety alignment with minimal parameter updates.

Contribution

SAILS leverages SAE-based disentanglement to construct an interpretable safety subspace, improving safety performance and interpretability over prior low-rank adaptation methods.

Findings

SAILS achieves up to 99.6% safety rate on Gemma-2-9B.

SAILS outperforms full fine-tuning by 7.4 points.

SAILS matches RLHF-based models' safety performance.

Abstract

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.