Do Sparse Autoencoders Learn Meaningful Concept Hierarchies?

arXiv:2606.22994 · cs.LG · Submitted 2026-06-22 · Read on arXiv

cs.LG

Submitted: 2026-06-22

Updated: 2026-10-06

Code: https://github.com/nlsgrndn/eval-HSAEs

Terminology

• Generalization — 1× in this paper · explained in 13 episodes
• Sparse Autoencoders (SAEs) — 1× in this paper · explained in 3 episodes

Sources

• Learning Multi-Level Features with Matryoshka Sparse Autoencoders
• A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoders
• Feature Hedging: Correlated Features Break Narrow Sparse Autoencoders
• Further Generalizations of the Jaccard Index
• From Flat to Hierarchical: Extracting Sparse Representations with Matching Pursuit
• Sparse Autoencoders Find Highly Interpretable Features in Language Models
• Sparse Autoencoders are Topic Models
• Interpretable Embeddings with Sparse Autoencoders: A Data Analysis Toolkit
• SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability
• Sparse Autoencoders Do Not Find Canonical Units of Analysis
• Controllable Text Generation for Large Language Models: A Survey
• From Atoms to Trees: Building a Structured Feature Forest with Hierarchical Sparse Autoencoders
• Incorporating Hierarchical Semantics in Sparse Autoencoder Architectures
• DINOv2: Learning Robust Visual Features without Supervision
• Sparse Autoencoders Learn Monosemantic Features in Vision-Language Models
• Automatically Interpreting Millions of Features in Large Language Models
• A Survey on Sparse Autoencoders: Interpreting the Internal Mechanisms of Large Language Models
• Eliminating Information Leakage in Hard Concept Bottleneck Models with Supervised, Hierarchical Concept Learning
• Gemma 4 Technical Report
• CFM: Language-aligned Concept Foundation Model for Vision

Related papers

• Polynomial-Augmented Neural Networks (PANNs) with Weak Orthogonality Constraints for Enhanced Function and PDE Approximation
• AIRL-S: Unifying Reinforcement Learning and Search-Based Test-Time Scaling via Adversarial Inverse Reinforcement Learning
• Transformers as Bayesian In-Context Experimenters: Smoothness-Adaptive Efficient ATE Estimation
• Convergence issues in Relational Concept Analysis based on AOC-posets
• Beliefs Beyond Posteriors: The Predictive Cost of ELBO Binding
• Understanding Diffusion Models via Ratio-Based Function Approximation with SignReLU Networks