Feature Superposition

Follow this topic Ask about this topic

Feature superposition studies how networks represent many features in a limited number of activation dimensions, often by sharing directions when features are sparse. This matters because the resulting interference can affect what models retain, compute, and reveal to interpretability tools. Researchers investigate it with synthetic sparse-feature models, geometric and compressed-sensing theory, and sparse autoencoders applied to pretrained language and vision networks. A central open question is whether good reconstruction and geometric recovery show that extracted features are genuinely used by the network, which requires causal tests such as ablation and steering.

Latest papers

24 papers3 papers in the last four weeks, down 40% on the four weeks before. 0.0% of all new papers.

Sparse Transformer: a Controlled Testbed for Mechanistic Interpretability
Oct 14, 2026Anton Korznikov, Alexey Dontsov, Elena TutubalinaMechanistic Interpretability Neuron-Level Interpretability
The Ball and the Box: Two Geometries of Computation in Superposition
Oct 8, 2026Xiaoyu Li, Lequan Lin, Dai Shi +3Feature Superposition Neural Representation Geometry
University of New South Wales · University of Sydney · University of Cambridge
Emergent phases of superposition: from partial to full representation
Sep 29, 2026Lihao Guo, Yizhou Liu, Jeff GorePhase Transitions in Neural Networks Feature Superposition
Massachusetts Institute of Technology Cambridge, MA 02139, USA
Hidden not Deleted: How Networks Suppress Entangled Features
Sep 23, 2026Akash Samanta, Manish Pratap Singh, Debasis ChaudhuriFeature Superposition Mechanistic Interpretability
Techno India University Salt Lake, Kolkata, India · DRDO Young Scientist Laboratory - CT Chennai, India
Topographic Training Concentrates Causal Circuits Without Improving Neuron Monosemanticity
Sep 21, 2026Gautam Ranka, Shubham Santosh Pandere, Aiden DsouzaFeature Superposition Vision Transformer
IvLabs, VNIT Nagpur, India.
A Dominant Diffuse Phase in the Sparse Autoencoder Phase Diagram
Sep 9, 2026Alexis D. PlascenciaFeature Superposition Representation Disentanglement
High-probability guarantees for linear accessibility in feature superposition
Sep 9, 2026Enrico VompaFeature Superposition Neural Representation Geometry
Applied Artificial Intelligence Group Tallinn University of Technology, Estonia
Feature Superposition in Neural Networks: From Theory to Practice
Sep 6, 2026Dai Shi, Xiaoyu Li, Andi Han +1Feature Superposition Neural Network Interpretability
University of Cambridge. · University of New South Wales. · University of Sydney.
Towards a mathematical theory of superposition
Aug 27, 2026Michael I. Ivanitskiy, John Jasper, Emily J. King +1Feature Superposition Sparse Recovery
Department of Applied Mathematics and Statistics, Colorado School of Mines, Golden, CO, USA · Department of Mathematics & Statistics, Air Force Institute of Technology, Wright-Patterson AFB, OH, USA · Department of Mathematics, Colorado State University, Fort Collins, CO, USA +2
Why Does Robustness Reduce Superposition?
Aug 23, 2026Adam ElimadiFeature Superposition Adversarial Examples
Independent Researcher
Scaling Interpretable Transformers with Parity Bottleneck Layers
Jul 22, 2026Andrew Mack, Kraig Yuheng Tou, Mark Henry +2Feature Superposition Transformer Interpretability
Principles of Intelligence · Independent
From Geometric Recovery to Causal Validation: A Reproducible Audit of Sparse Autoencoder Features, from Superposition Geometry to Causal Inertness
Jul 13, 2026Mohamed Abdessalem BalFeature Superposition Neural Representation Geometry
Independent Researcher, Algiers, Algeria
Resolving superposition in AI for interpretability and cross-modal alignment in patient-neuronal images
Jun 30, 2026Jisung Park, Seohyeon Kang, Daeun Yoo +8Feature Superposition Neural Representation Geometry
1KAIST · 2Konyang University · 4UCL Queen Square Institute of Neurology & The Francis Crick Institute +1
Sparsity, Superposition, and Forgetting: A Mechanistic Study of Representation Retention in Continual Learning
Jun 18, 2026Jan Wasilewski, Jędrzej Kozal, Michał Woźniak +1Feature Superposition Representation Learning
Rochester Institute of Technology · Rochester NY, USA · Wrocław University of Science and Technology +1
Effects of sparsity and superposition on loss in simple autoencoders
Jun 16, 2026Mriganka Basu Roy Chowdhury, Eric McLaughlin WeinerFeature Superposition Activation Sparsity
Department of Statistics, UC Berkeley · Department of Materials Science, UC Berkeley
Representational Capacity: Geometric Limits on Feature Representation in Transformer Language Models
Jun 1, 2026Alexander GuhaFeature Superposition Transformer Expressivity
Arizona State University
Probing for Representation Manifolds in Superposition
May 18, 2026Alexander ModellFeature Superposition Neural Representation Geometry
Department of Mathematics Imperial College London, U.K.
Linear-Readout Floors and Threshold Recovery in Computation in Superposition
May 2, 2026Hector Borobia, Elies Seguí-Mas, Guillermina Tormo-CarbóFeature Superposition
VRAIN – Valencian Research Institute for Artificial Intelligence, Universitat Politècnica de València, Valencia, Spain · Department of Economics and Social Sciences, Universitat Politècnica de València, Valencia, Spain · Department of Business Organisation, Universitat Politècnica de València, Valencia, Spain
Adjoint Inversion Reveals Holographic Superposition and Destructive Interference in CNN Classifiers
Apr 30, 2026Kaixiang ShuFeature Superposition Neural Network Interpretability
Independent Researcher Chongqing, China
Structural Instability of Feature Composition
Apr 18, 2026Yunpeng ZhouFeature Superposition Transformer Interpretability
Whiteknights House, Whiteknights, Reading, RG6 6UR
Understanding Emergent Misalignment via Feature Superposition Geometry
Apr 7, 2026Gouki Minegishi, Hiroki Furuta, Takeshi Kojima +2Emergent Misalignment in Language Models LLM Alignment
The University of Tokyo · Google DeepMind
Polysemanticity or Polysemy? Lexical Identity Confounds Superposition Metrics
Apr 1, 2026Iyad Ait Hou, Rebecca HwaWord Sense Disambiguation Feature Superposition
Department of Computer Science George Washington University Washington, D.C, USA
Similarity of Neural Network Representations in Superposition
Mar 31, 2026Sunny Liu, Habon Issa, André Longon +4Feature Superposition Representational Similarity Analysis
Cold Spring Harbor Laboratory · UC San Diego · Anthropic +2
From Data Statistics to Feature Geometry: How Correlations Shape Superposition
Mar 10, 2026Lucas Prieto, Edward Stevinson, Melih Barsbey +2Feature Superposition Neural Network Interpretability

Oct 14, 2026J/K move · Enter open · S save

Sparse Transformer: a Controlled Testbed for Mechanistic Interpretability

Anton Korznikov, Alexey Dontsov, Elena Tutubalina

Mechanistic Interpretability Neuron-Level Interpretability Transformer Sparse Autoencoders Feature Superposition

Open paper AskPDF arXiv

Mechanistic interpretability aims to decompose neural network computations into human-understandable features, yet progress is limited by the absence of ground-truth units to evaluate against. We address this with a Sparse Transformer (ST): a 5-layer language model trained on TinyStories whose wide, TopK-sparse MLP forces neuron specialization. We show that ST neurons are highly interpretable: they outperform a loss-matched dense transformer by 43% on AutoInterp, and circuits of a median of 12 neurons suffice to reproduce individual model predictions. This testbed allows us to directly compare feature learning methods: Transcoder, which approximates the MLP input-output mapping, recovers ST's neuron vectors with 52.6% higher cosine similarity than Sparse Autoencoder, which only reconstructs outputs. Furthermore, the ST enables a controlled study of polysemanticity: distilling its sparse MLP into a dense counterpart reveals how compression leads to feature superposition. These results show how a known reference basis sharpens evaluation and understanding, a role we hope the ST can serve for the field.