The paper proposes a new unsupervised safety detection method for large language models that relies on anomaly detection rather than supervised training on unsafe data. By leveraging local sparsity in a linear representation space obtained via a sparse autoencoder, the authors develop a framework for locally masked SAE-based anomaly detection, providing theoretical support and empirical validation across multiple architectures and datasets. When calibrated with only 1% out-of-distribution data, the method achieves near‑optimal performance while using just 1–2% of SAE neurons for computation.
By Xin Chen, Gil Kur, Alexander Shevchenko, Andreas Krause
The paper proposes a novel unsupervised safety detection method for large language models that relies on local sparsity in a linear representation space recovered via a sparse autoencoder. By masking SAE neurons based on shared active support among nearby points, the authors develop a locally masked anomaly detection framework with theoretical backing. Experiments across multiple architectures and datasets—including capability‑testing and safety‑specific sets—show that using only 1–2% of SAE neurons and a small amount of out‑of‑distribution data yields near‑optimal safety detection performance.
The paper introduces a controlled auditing framework using parametric CAD to separate failures in encoding, expression, and activation-level control in frozen large language models (LLMs). Applied to eight pretrained models and random controls, it finds that pretraining mainly improves linear decoding of local relations, while global constraint status is already highly decodable from random representations. Despite strong local decodability, the models’ zero‑shot predictions lag far behind supervised probe performance, and simple activation steering fails to provide reliable control; however, low‑rank fine‑tuning with only 400 labeled relation pairs can substantially improve generation on three models, indicating that the representation–behavior gap can be narrowed through adaptation.
By Man Liang, Xinzhao Cheng, Faizan Wajid
The paper introduces Dynamic DAE Guardrails (DSG), a method that uses Dynamic Sparse Autoencoders to perform precision unlearning in large language models. DSG leverages principled feature selection and a dynamic classifier to target activation-based unlearning, outperforming existing gradient‑based methods in terms of computational efficiency, stability, sequential unlearning, resistance to relearning attacks, data efficiency, and interpretability.
By Aashiq Muhamed, Jacopo Bonato, Mona Diab, Virginia Smith
JEPAMatch introduces a new semi‑supervised learning framework that replaces traditional output‑thresholding with explicit geometric shaping of latent representations. By combining the FlexMatch loss with a latent‑space regularization inspired by LeJEPA, the method encourages isotropic Gaussian structure in the embedding space, mitigating class imbalance and noisy pseudo‑labels. Experiments on CIFAR‑100, STL‑10, and Tiny‑ImageNet show consistent performance gains and faster convergence compared to existing FixMatch‑based baselines.
By Ali Aghababaei-Harandi, Aude Sportisse, Massih-Reza Amini
arXiv:2606. 07007v1 Announce Type: cross Abstract: We propose a unified mathematical framework for a geometric understanding of concept learning and neuron interpretation in sparse autoencoders (SAEs).
By Chenhao Zhang, Chris Lin, Su-In Lee
arXiv:2606. 06333v1 Announce Type: new Abstract: Sparse Autoencoders (SAEs) are widely used for mechanistic interpretability in large language models, yet their formulation assigns each latent feature a single decoder direction, implicitly assuming features to be one-dimensional.
By Seyed Arshan Dalili, Mehrdad Mahdavi
arXiv:2512. 07355v2 Announce Type: replace Abstract: Two traditions of interpretability have evolved side by side but seldom spoken to each other: Concept Bottleneck Models (CBMs), which prescribe what a concept should be, and Sparse Autoencoders (SAEs), which discover what concepts emerge.
By Alexandre Rocchi, Thomas Fel, Gianni Franchi
DIFFINT is a reconstruction‑based anomaly detector that uses a differentiable autoencoder with a latent bottleneck composed of soft, axis‑aligned interval memberships. Each latent unit represents a human‑readable hyper‑rectangle in feature space, allowing the model to encode how strongly an instance falls inside each interval and to compute reconstruction error as the anomaly score. The method provides a certified lower bound on reconstruction error for points outside all active intervals, a suppression mechanism for sparse abnormalities, and a closed‑form, label‑free importance ranking for each (unit, feature) pair, achieving top performance on 48 ADBench benchmarks against 22 baselines.
By Lamine Diop, Marc Plantevit
The paper demonstrates that the common-geodesic condition—where every output triple shares a geodesic point in a metric—does not ensure Fisher consistency for the structured SVM with the standard coordinate-wise argmax decoder. It presents minimal counterexamples, including a four-output star and a tree-metric classification, showing that only path-shaped trees maintain argmax consistency. The study also identifies the smallest full-support counterexamples and provides exact primal-dual certificates for all optimality claims.
By Jintao Fei, Jiangying Luo
arXiv:2606. 13803v1 Announce Type: new Abstract: Enforcing functional inequality constraints such as monotonicity and convexity in neural networks is a fundamental challenge in many industrial and scientific applications.
By Ruben Wiedemann, Antoine Jacquier, Lukas Gonon
The paper introduces PolyFormer, a physics-informed machine learning framework that learns compact polytopic representations of complex constraints. By transforming constraint-induced geometry into efficient polytopic reformulations, PolyFormer reduces optimization complexity and enables the use of standard solvers. Evaluations on large‑scale resource aggregation, network‑constrained optimization, and uncertainty‑aware optimization show up to 6,400‑fold speedups and 99.87% memory savings while keeping feasibility and objective errors low.
By Yilin Wen, Yi Guo, Bo Zhao, Wei Qi, Zechun Hu, Colin Jones, Jian Sun