Large Language Models (LLMs) remain vulnerable to optimization-based jailbreak attacks that exploit internal gradient structure. While Sparse Autoencoders (SAEs) are widely used for interpretability, their robustness implications remain underexplored. We present a study of integrating pretrained SAEs into transformer residual streams at inference time, without modifying model weights or blocking gradients. Across four model families (Gemma, LLaMA, Mistral, Qwen) and two strong white-box attacks (GCG, BEAST) plus three black-box benchmarks, SAE-augmented models achieve up to a 5x reduction in jailbreak success rate relative to the undefended baseline and reduce cross-model attack transferability. Parametric ablations reveal (i) a monotonic dose-response relationship between L0 sparsity and attack success rate, and (ii) a layer-dependent defense-utility tradeoff, where intermediate layers balance robustness and clean performance. These findings are consistent with a representational bottleneck hypothesis: sparse projection reshapes the optimization geometry exploited by jailbreak attacks.
Zhejiang University(浙江大学);Hongkong Polytechnic University(香港理工大学);University College London(伦敦大学学院);Westlake University(西湖大学);King's College London(国王学院);Southern University of Science and Technology(南方科技大学)
Graph-Regularized Sparse Autoencoders for LLM Safety Steering
ELLIS Institute Tübingen(图宾根ELLIS研究所);Max Planck Institute for Intelligent Systems(智能系统马克斯·普朗克研究所);Intesa Sanpaolo(Intesa Sanpaolo公司);University of Southern California(南加州大学)
SplInterp: Improving our Understanding and Training of Sparse Autoencoders
School of Mathematics University of Birmingham(数学学院英国伯明翰大学);Ideami Studios(Ideami工作室);Department of Mathematics and Statistics University of Limerick(数学与统计学学院英国利默里克大学);XRAI Inc.(XRAI公司);Department of Computer Science Brown University(计算机科学系布朗大学)
CRISP: Persistent Concept Unlearning via Sparse Autoencoders
Technion – Israel Institute of Technology(技术ion-以色列理工学院);Boston University(波士顿大学);University of Zagreb, FER(Zagreb大学,FER);Kempner Institute, Harvard University(哈佛大学 Kempner研究所)
Understanding Refusal in Language Models with Sparse Autoencoders
Nanyang Technological University(南洋理工大学);Singapore University of Technology and Design(新加坡科技设计大学);Digital Trust Centre(数字信任中心);Institute of High Performance Computing (IHPC)(高性能计算研究所)
SAVE: Sparse Autoencoder-Driven Visual Information Enhancement for Mitigating Object Hallucination
Department of Electrical and Computer Engineering, Seoul National University(电子与计算机工程系,首尔国立大学);Daegu Gyeongbuk Institute of Science and Technology(大邱庆州科学技术院);IPAI, AIIS, ASRI, INMC, and ISRC, Seoul National University(IPAI、AIIS、ASRI、INMC 和 ISRC,首尔国立大学)
SAE-RNA: A Sparse Autoencoder Model for Interpreting RNA Language Model Representations
Department of Computer Science, University of California, Berkeley(加州大学伯克利分校计算机科学系);Department of Development Engineering, University of California, Berkeley(加州大学伯克利分校发展工程系)
Fully Bayesian Autoencoders with Latent Sparse Gaussian Processes
EURECOM(欧洲通信研究院);University of California, Irvine(加州大学欧文分校)
自动编码器及其变体是表示学习和生成建模中使用最广泛的模型之一。然而,基于自动编码器的模型通常假设学习的表示是i.i.d.。并且不能捕获数据样本之间的相关性。为了解决这个问题,我们提出了一种新的稀疏高斯过程贝叶斯自动编码器(SGPAE)模型,在该模型中,我们对贝叶斯自动编码器的潜在空间施加完全贝叶斯稀疏高斯过程先验。我们通过随机梯度Hamilton Monte Carlo对该模型进行后验估计。我们在广泛的表示学习和生成建模任务上定性和定量地评估了我们的方法,并表明我们的方法始终优于依赖于变分自编码器的多个替代方案。
Temporal Sparse Autoencoders: Leveraging the Sequential Nature of Language for Interpretability
Harvard School of Engineering and Applied Science(哈佛工程与应用科学学院);Kempner Institute for the Study of Natural & Artificial Intelligence(自然与人工智能研究学院);Harvard Business School(哈佛商学院)