Sparse Autoencoders (SAEs) are widely used to interpret neural networks by identifying meaningful concepts from their representations. However, do SAEs truly uncover all concepts a model relies on, or are they inherently biased toward certain kinds of concepts? We introduce a unified framework that recasts SAEs as solutions to a bilevel optimization problem, revealing a fundamental challenge: each SAE imposes structural assumptions about how concepts are encoded in model representations, which in turn shapes what it can and cannot detect. This means different SAEs are not interchangeable -- switching architectures can expose entirely new concepts or obscure existing ones. To systematically probe this effect, we evaluate SAEs across a spectrum of settings: from controlled toy models that isolate key variables, to semi-synthetic experiments on real model activations and finally to large-scale, naturalistic datasets. Across this progression, we examine two fundamental properties that real-world concepts often exhibit: heterogeneity in intrinsic dimensionality (some concepts are inherently low-dimensional, others are not) and nonlinear separability. We show that SAEs fail to recover concepts when these properties are ignored, and we design a new SAE that explicitly incorporates both, enabling the discovery of previously hidden concepts and reinforcing our theoretical insights. Our findings challenge the idea of a universal SAE and underscores the need for architecture-specific choices in model interpretability. Overall, we argue an SAE does not just reveal concepts -- it determines what can be seen at all.

Sparse autoencoders (SAEs) are used to disentangle the dense, polysemantic internal representations of large language models (LLMs) into interpretable, monosemantic concepts. However, standard $\ell_1$-regularized SAEs suffer from feature starvation (dead neurons) and shrinkage bias, often requiring computationally expensive heuristic resampling and nondifferentiable hard-masking methods to bypass these challenges. We argue that feature starvation is not merely an empirical artifact of poor data diversity, but a fundamental optimization-geometric pathology of overcomplete dictionaries: the $\ell_1$-induced sparse coding map is unstable and fundamentally misaligned with shallow, amortized encoders. To address this structural instability, we introduce adaptive elastic net SAEs (AEN-SAEs), a fully differentiable architecture grounded in classical sparse regression. AEN-SAEs combine an $\ell_2$ structural term that enforces strong convexity and Lipschitz stability with adaptive $\ell_1$ reweighting that eliminates shrinkage bias and suppresses spurious features, thereby jointly controlling the curvature and interaction structure of the induced polyhedral geometry. Theoretically, we show that AEN-SAEs yield a Lipschitz-continuous sparse coding map and recover the global feature support under mild assumptions. Empirically, across synthetic settings and LLMs (Pythia 70M, Llama 3.1 8B), AEN-SAEs mitigate feature starvation without auxiliary heuristics while maintaining competitive reconstruction abilities.

Techniques for concept extraction, such as sparse autoencoders and transcoders, aim to extract high-level symbolic concepts from low-level nonsymbolic representations. When these extracted concepts are used for downstream tasks such as model steering and unlearning, it is essential to understand their guarantees, or lack thereof. In this work, we present a unified theoretical framework for unsupervised concept extraction, in which we frame the task of concept extraction as identifying a generative model. We present a general meta-theorem for identifiability, which reduces the problem of establishing identifiability guarantees to the problem of characterizing the intersection of two sets. As we demonstrate on a range of widely-used approaches, this meta-theorem substantially simplifies the task of proving such guarantees, thus paving the way for the development of new, principled approaches for concept extraction.

By studying large language models as if they were living things instead of computer programs, scientists are discovering some of their secrets for the first time. How large is a large language model? Think about it this way. In the center of San Francisco there's a hill called Twin Peaks from which you can view nearly the entire city. Picture all of it--every block and intersection, every neighborhood and park, as far as you can see--covered in sheets of paper. Now picture that paper filled with numbers. LLMs contain a LOT of parameters. That's one way to visualize a large language model, or at least a medium-size one: Printed out in 14-point type, a 200-billion-parameter model, such as GPT4o (released by OpenAI in 2024), could fill 46 square miles of paper--roughly enough to cover San Francisco.

Transformers generate valid and diverse chemical structures, but little is known about the mechanisms that enable these models to capture the rules of molecular representation. We present a mechanistic analysis of autoregressive transformers trained on drug-like small molecules to reveal the computational structure underlying their capabilities across multiple levels of abstraction. We identify computational patterns consistent with low-level syntactic parsing and more abstract chemical validity constraints. Using sparse autoencoders (SAEs), we extract feature dictionaries associated with chemically relevant activation patterns. We validate our findings on downstream tasks and find that mechanistic insights can translate to predictive performance in various practical settings.

Interpreto is a Python library for post-hoc explainability of text HuggingFace models, from early BERT variants to LLMs. It provides two complementary families of methods: attributions and concept-based explanations. The library connects recent research to practical tooling for data scientists, aiming to make explanations accessible to end users. It includes documentation, examples, and tutorials. Interpreto supports both classification and generation models through a unified API. A key differentiator is its concept-based functionality, which goes beyond feature-level attributions and is uncommon in existing libraries. The library is open source; install via pip install interpreto. Code and documentation are available at https://github.com/FOR-sight-ai/interpreto.

Since the advent of machine learning, interpretability has remained a persistent challenge, becoming increasingly urgent as generative models support high-stakes applications in drug and material discovery. Recent advances in large language model (LLM) architectures have yielded chemistry language models (CLMs) with impressive capabilities in molecular property prediction and molecular generation. However, how these models internally represent chemical knowledge remains poorly understood. In this work, we extend sparse autoencoder techniques to uncover and examine interpretable features within CLMs. Applying our methodology to the Foundation Models for Materials (FM4M) SMI-TED chemistry foundation model, we extract semantically meaningful latent features and analyse their activation patterns across diverse molecular datasets. Our findings reveal that these models encode a rich landscape of chemical concepts. We identify correlations between specific latent features and distinct domains of chemical knowledge, including structural motifs, physicochemical properties, and pharmacological drug classes. Our approach provides a generalisable framework for uncovering latent knowledge in chemistry-focused AI systems. This work has implications for both foundational understanding and practical deployment; with the potential to accelerate computational chemistry research.

Large language models (LLMs) face critical safety challenges, as they can be manipulated to generate harmful content through adversarial prompts and jailbreak attacks. Many defenses are typically either black-box guardrails that filter outputs, or internals-based methods that steer hidden activations by operationalizing safety as a single latent feature or dimension. While effective for simple concepts, this assumption is limiting, as recent evidence shows that abstract concepts such as refusal and temporality are distributed across multiple features rather than isolated in one. To address this limitation, we introduce Graph-Regularized Sparse Autoencoders (GSAEs), which extends SAEs with a Laplacian smoothness penalty on the neuron co-activation graph. Unlike standard SAEs that assign each concept to a single latent feature, GSAEs recover smooth, distributed safety representations as coherent patterns spanning multiple features. We empirically demonstrate that GSAE enables effective runtime safety steering, assembling features into a weighted set of safety-relevant directions and controlling them with a two-stage gating mechanism that activates interventions only when harmful prompts or continuations are detected during generation. This approach enforces refusals adaptively while preserving utility on benign queries. Across safety and QA benchmarks, GSAE steering achieves an average 82% selective refusal rate, substantially outperforming standard SAE steering (42%), while maintaining strong task accuracy (70% on TriviaQA, 65% on TruthfulQA, 74% on GSM8K). Robustness experiments further show generalization across LLaMA-3, Mistral, Qwen, and Phi families and resilience against jailbreak attacks (GCG, AutoDAN), consistently maintaining >= 90% refusal of harmful content.