CS Paper Radar

Exploring Text Classification Models with Sparse Autoencoders

Daniel Kerrigan, Brian Barr, Enrico Bertini

cs.HC ·submitted 2026-09-17 23:14 UTC · v1
published

Abstract

As language models (LMs) rise in prominence, there is interest in making them more transparent in order to better understand their internal behavior. Recent interpretability work has focused on using sparse autoencoders (SAEs) to break down neuron activations at a given layer in the LM into human-understandable features, where each feature represents a concept that the model has learned. In this paper, we share work on using SAEs to analyze the behavior of text classification LMs. We present techniques for exploring the relationships between the SAE's features and the model's predictions and errors. We integrate these techniques into SAEfarer, a tool for analyzing concepts learned by text classification LMs. We assess SAEfarer in an expert pilot evaluation with five Ph.D. students.

Comment
11 pages, 13 figures. Accepted as a short paper at IEEE VIS 2026

Closest papers

Nearest neighbours by meaning across the whole index.

Building Auditable AI Text Detectors with Sparse Autoencoders

cs.CL cs.AI ·2026-10-21 ·Liam Dugan, Callum McDougall, Matthew Denton, Philip Pham, Christine Kaeser-Chen +1 more
top venue similarity 0.72
Abstract · EMNLP 2026

AI-generated text detectors are increasingly deployed in consequential settings---peer review, education, admissions, journalism, and content moderation---where their outputs directly affect individuals. Yet the strongest detectors are fine-tuned neural classifiers whose decisions are opaque, while transparent alternatives (stylometric and linguistic-feature classifiers) underperform them by wide margins. We present a pipeline that converts a fine-tuned neural detector into a classifier that operates on human-readable concepts. We accomplish this by training sparse autoencoders (SAEs) directly on a detector's residual stream, assigning auto-interpretation labels to each SAE feature, and retaining only features whose generated descriptions pass a global fidelity threshold. We then train a lightweight classifier on just the interpretable features. We find that our SAEs recover latent concepts that correspond to highly specific and predictive linguistic "tells" of AI-generated text. Classifiers built on these extracted features are not only highly interpretable but exhibit stronger performance on out-of-distribution texts than either neural or stylometric alternatives.

0.50

SAEVerbalizer: Generating Explanations for Sparse Autoencoder Features via Representation Verbalization

cs.CL ·2026-08-13 ·Weihan Meng, Hongzhu Guo, Yi Jing, Dewen Liu, Zijun Yao +3 more
similarity 0.72
Abstract

Sparse autoencoders (SAEs) are proposed to extract numerous features from large language model (LLM) representations, yet explaining these features still relies primarily on external observation. This reliance leads to superficial explanations inferred from observed model behavior and computational inefficiency from collecting such behavioral evidence at scale. We introduce SAEVerbalizer, a framework that injects SAE decoder directions into an LLM's representations and fine-tunes the LLM's downstream layers to generate natural-language explanations of the injected features. Once trained, the resulting verbalizer explains SAE features directly from decoder directions, addressing both limitations. Our experiments show that the learned verbalization capability generalizes to unseen features, transfers across separately trained SAE dictionaries, and, with a lightweight adapter, extends to SAE features from different LLMs. Intervention experiments show that injecting multiple directions yields an explanation combining their meanings, while reversing individual directions produces corresponding meaning shifts.

0.39

Sparse Autoencoders for Mechanistic Interpretability in NLP: A Critical Survey of Recoverability, Interpretability, and Utility

cs.CL ·2026-10-14 ·Salima Lamsiyah, Adelaide Danilov, Oleksandr Marchenko Breneur, Aria Nourbakhsh
survey similarity 0.71
Abstract · 28 pages, 5 figures

Sparse autoencoders (SAEs) have emerged as a promising approach to mechanistic interpretability by decomposing language-model activations into sparse, potentially interpretable features. However, differences in design assumptions, evaluation objectives, and experimental settings complicate conclusions about their effectiveness. We present a critical survey of 138 studies on SAEs in natural language processing. Our contributions are fourfold: (1) we propose a task-grounded taxonomy connecting dictionary design, feature operations, and NLP functions; (2) we examine how sparsity allocation, geometry, data coverage, and inference affect feature recoverability; (3) we assess evidence for semantic interpretability, mechanistic faithfulness, and behavioral utility; and (4) we synthesize findings across multilingual processing, retrieval, reasoning, classification, and behavioral regulation, highlighting the roles of objectives, baselines, supervision, and intervention pipelines. Through comparative analyses of published results, we reconcile apparent disagreements, identify unresolved questions, and propose directions for more rigorous evaluation. An anonymous research register https://anonymous.4open.science/r/sae-research-register-C331/register.html accompanies the survey.

0.22

Gradients for Interventions and Activations for Detection: Targeted Feature Learning in Language Models

cs.CL cs.LG ·2026-09-26 ·Jonathan Drechsel, Steffen Herbold
similarity 0.68
Abstract

Model-internal features can be studied through both their ability to identify a specified concept and their causal effect when manipulated, e.g., through steering or weight editing. A prominent approach to feature learning is Sparse Autoencoders (SAEs), which learn broad feature dictionaries whose relation to particular concepts is typically identified post hoc. However, many interpretability questions are instead hypothesis-driven and concern a concept specified in advance. We study this setting as targeted feature learning, where a single feature is constructed for such a predefined concept. We present a controlled comparison across three model signals (activation values, activation gradients, and parameter gradients) and two estimators (contrastive mean and a learned one-dimensional encoder-decoder), yielding six targeted methods, with CAA and GRADIEND as existing instances and four new methods covering the remaining combinations. We compare these methods against pretrained SAEs across 15 tasks and three language models, evaluating both detection and causal intervention. Across models, the strongest detection performance is achieved by contrastive activation value methods, whereas the strongest intervention performance is achieved by gradient-based methods. Overall, our results show that targeted feature quality depends jointly on the model signal and estimator, with detection and intervention capturing complementary properties.

0.12

LLM Layers Immediately Correct Each Other

cs.CL cs.LG ·2026-09-07 ·Arjun Patrawala, Jiahai Feng, Erik Jones, Jacob Steinhardt
top venuecode/datahighly cited similarity 0.67
Abstract · Published at NeurIPS 2025

Recent methods in language model interpretability employ techniques such as sparse autoencoders to decompose residual stream contributions into linear, semantically meaningful features. Such methods are commonly interpreted as identifying features that persist in the residual stream and that subsequent layers build upon. We challenge this view by identifying the Transformer Layer Correction Mechanism (TLCM), wherein adjacent transformer layers systematically counteract portions of each other's contributions. TLCM appears in 5 out of 7 major open-source model families and activates across nearly all tokens in diverse texts. We show that TLCM emerges during pretraining, operates most strongly on contextually dependent tokens, and adaptively calibrates its correction strength based on the preceding layer's output. Using the layer Jacobian, we further show that TLCM selectively corrects specific subspaces while reinforcing others, which we interpret through a ``propose-and-reject'' framework in which layers propose candidate features and subsequent layers selectively remove inappropriate ones. This dynamic suggests that the residual stream at any layer contains transient proposals alongside persistent features, helping explain why SAE feature descriptions often have low specificity, why effective model steering requires extreme feature amplification, and why transcoders hold a theoretical advantage over SAEs.

0.72

Interpreting Object-Dependent Concept Brittleness in Text-to-Image Diffusion Models

cs.CV cs.MM ·2026-09-09 ·Yifan Yuan, Xiangyu Liu, Hongming Shan, Yu Han, Yu Jiang +3 more
top venuecode/data similarity 0.67
Abstract · Accepted at ACM MM 2026. 27 pages, 17 figures, including appendices

Although text-to-image diffusion models generally exhibit strong prompt-following ability, we identify a persistent and previously underexplored failure pattern in which a small subset of prompts differing only in the object consistently fails to realize the same target concept under identical generation settings. We term this phenomenon object-dependent concept brittleness. Such cases suggest systematic internal blind spots rather than random sampling noise. In this paper, we present an interpretability-oriented framework to audit and minimally correct these failures. Our key idea is to analyze denoising trajectories in a step-wise sparse autoencoder (SAE) space, where abstract style and attribute concepts become more separable than in the raw denoising representation. This sparse space enables us to compare successful and failed generations, identify concept dimensions whose evidence is missing, weakened, or temporally delayed, and construct class-level concept prototypes from reliable class-consistent samples. Based on this audit process, we introduce a lightweight inference-time correction strategy that interpolates denoising features toward the corresponding prototype in SAE space. Rather than serving as a task-specific retraining method, this intervention acts as a validation of the diagnosed concept deficiency. We evaluate the proposed framework on style and attribute failure cases across multiple diffusion backbones, with significant improvements in concept consistency, text fidelity, and repair success. Further analyses show that deeper denoising representations provide clearer concept structure, while early-stage intervention offers the strongest correction leverage. Code is available at https://github.com/Metecade/Object-Dependent-Concept-Brittleness.

0.46

Making Clinical Language Models Auditable: Concept-Guided Fine-Tuning for Robust Prediction

cs.CL cs.AI ·2026-08-27 ·Jin Mu, Guanhua Chen
top venue similarity 0.67
Abstract

Clinical language models can achieve strong in-hospital accuracy yet fail under deployment shifts because they exploit note-specific artifacts (e.g., templates, separators, boilerplate) that do not reflect patient state. We propose CAST (Concept-guided Artifact Suppression Tuning), an SAE-based framework for auditable clinical text classification. CAST uses Sparse Autoencoders to expose sparse, human-auditable features from intermediate Transformer activations, labels SAE latents with an LLM-assisted interpretation pipeline and ICD-10 retrieval constraints, suppresses verified artifact latents via residual subtraction during fine-tuning, and provides post-hoc per-concept attributions for auditing model decisions. On MIMIC-IV discharge-note mortality prediction, CAST improves over its corresponding fine-tuned encoder baselines and remains competitive with strong LLM baselines, while producing a feature-level audit trail of the clinical concepts that support each prediction and the artifact concepts suppressed during training.

0.44

When Pruning Meets Interpretability: Preserving Sparse Autoencoder Robustness in LLMs

cs.LG ·2026-08-26 ·Suchit Gupte, Xueru Zhang, Mohammad Mahdi Khalili
code/data similarity 0.66
Abstract

Sparse autoencoders (SAEs) are widely used to interpret the internal representations of large language models (LLMs), yet their reliability under post-hoc model compression remains poorly understood. We present a systematic study of how pruning affects SAE behavior and theoretically show that, for a fixed SAE, its impact is governed by perturbation energy, a covariance-weighted norm. This perspective exposes a key limitation of magnitude pruning: by ignoring activation geometry, it distorts the learned representation space and degrades SAE functionality. Activation-aware methods such as Wanda and SparseGPT, in contrast, implicitly control perturbation energy and are therefore substantially more robust at preserving SAE behavior. We further reveal a consistent structural vulnerability across all pruning methods: middle layers are significantly more sensitive to pruning than early or late layers. Guided by this insight, we propose a layer-wise sparsity allocation strategy, achieving lower perplexity under the same average pruning sparsity. Experiments across four model architectures validate our theoretical findings. Code is publicly available at https://github.com/osu-srml/sae-robustness-under-pruning/tree/main.

0.32