🤖 machine learning

Do Sparse Autoencoders Learn Meaningful Concept Hierarchies?

This paper proposes a rigorous evaluation protocol for assessing concept hierarchies in sparse autoencoders and reveals that while these models can form sensible hierarchies, their quality is systematically compromised by both hard and soft forms of feature absorption, highlighting a fundamental tension in current approaches.

Original authors: Nils Grandien, David Steinmann, Felix Friedrich, Kristian Kersting

Published 2026-06-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Nils Grandien, David Steinmann, Felix Friedrich, Kristian Kersting

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a giant, chaotic library inside a super-smart computer. This library contains millions of tiny "ideas" or "concepts" the computer has learned about the world, like "tennis racket," "pine cone," or "sport."

Currently, most of these libraries are just huge piles of books with no organization. If you want to find something specific, you have to dig through the whole mess. The computer might know what a "tennis racket" is, but it doesn't naturally understand that a racket is a type of sports equipment, or that "tennis" is a type of "sport."

This paper is about trying to organize that messy library into a neat, hierarchical structure (like a family tree for ideas) and figuring out how well current computer methods are actually doing at this task.

The Goal: Building a "Family Tree" for Ideas

Humans are great at organizing things. We know that a "Golden Retriever" is a kind of "Dog," and a "Dog" is a kind of "Animal." This is a hierarchy.

The researchers wanted to see if Sparse Autoencoders (SAEs)—a tool computers use to find hidden patterns in data—could learn to build these same kinds of family trees on their own, without being told what the categories are.

The Problem: The "Absorption" Glitch

The paper discovered a major snag in how these computers organize their thoughts. They call it Absorption.

Think of it like a loud child at a dinner table.

The Parent Concept: "Sport" (The parent).
The Child Concept: "Tennis Serve" (The child).

Ideally, if the computer sees a tennis serve, it should light up the "Tennis Serve" button and the "Sport" button, because a serve is a type of sport.

However, the researchers found that often, when the "Tennis Serve" button lights up brightly, the "Sport" button gets silenced or ignored. The specific idea "absorbs" the general idea.

Hard Absorption: The parent button turns off completely when the child is active.
Soft Absorption: The parent button turns on, but very weakly, while the child is very bright. It's like the parent is whispering while the child is shouting.

This is a problem because if the computer silences the "Sport" button when it sees a tennis serve, it loses the ability to understand the broader context.

The Solution: A New Rulebook and Scorecard

Since no one had agreed on how to measure if a computer's idea-tree was "good," the authors created a new rulebook. They defined what a good hierarchy should look like in three ways:

The Concepts Must Make Sense: Each "book" in the library must be a real, understandable idea (like "pine cone"), not gibberish.
The Structure Must Be Logical: A "child" concept must be more specific than its "parent." (e.g., "Tennis Racket" is more specific than "Sport").
The Activity Must Match the Logic: This is the big one. If a child is active, the parent must also be active and strong. If you see a "Tennis Racket," the computer should confidently say "Yes, that is a Sport" at the same time.

They then built a Scorecard (metrics) to grade different computer methods based on these rules.

What They Found

The researchers tested several different computer methods on images (like photos of sports and nature). Here is the verdict:

The Good News: The computers can find meaningful ideas. They can group things like "birds" and "roosters" together. The basic building blocks are there.
The Bad News: The organization is usually very shallow (only 1 or 2 levels deep) and the "Absorption" problem is everywhere.
- Most methods failed to keep the "parent" and "child" buttons active together.
- Some methods tried to force a hierarchy during training, but they ended up creating trees where the "children" were actually more abstract than the "parents," which is backwards.
- One method (ActMSAE) managed to build a deeper tree, but even it struggled with the "soft absorption" issue where the parent gets too quiet when the child is loud.

The Takeaway

The paper concludes that while computers are getting better at finding individual concepts, teaching them to organize those concepts into a logical, working family tree is still very hard. The main enemy is Absorption—the tendency for specific details to drown out the general categories they belong to.

To fix this, future computer models might need to stop being so "sparse" (picking only one or two ideas) and instead learn to activate both the specific detail and the general category at the same time, ensuring the "parent" never gets silenced by the "child."

Technical Summary: Do Sparse Autoencoders Learn Meaningful Concept Hierarchies?

Problem Statement
Sparse Autoencoders (SAEs) have emerged as a primary tool for unsupervised concept discovery in large language and vision-language models, decomposing model activations into high-dimensional, sparse latent features. However, as these feature spaces grow, they become difficult to navigate and interpret due to a lack of inherent structure. While humans naturally organize concepts hierarchically (e.g., tennis $\to$ tennis racket), current SAE approaches often produce unordered concept spaces. Although recent methods attempt to impose hierarchical structures—either explicitly during training or implicitly via constraints—there is no agreed-upon framework for what constitutes a "meaningful" feature hierarchy. Consequently, rigorous comparison between methods is difficult, and evaluation has largely relied on qualitative illustrations or fragmented quantitative protocols.

Methodology
To address the lack of standardized evaluation, the authors derive a comprehensive set of requirements for generalization/specialization hierarchies in unsupervised concept learning, drawing from semantic net research, taxonomy evaluation, and recent SAE literature. These requirements are categorized into three levels:

Concept-Level: Focuses on individual features.
- Interpretability (A1): Nodes must correspond to human-understandable concepts.
Structure-Level: Focuses on the static graph topology.
- Hierarchical Abstractness (B1): Child nodes must be semantically more specific than their parents.
- Sibling Similarity (B2): Children of the same parent should form a coherent, semantically related set.
Co-Activation-Level: Focuses on dynamic activation behavior on concrete inputs.
- Activation Implication (C1): An active child should imply an active parent (no "hard" absorption).
- Activation Dominance (C2): Parent activation strength should be at least as strong as the child's when both are active.
- Frequent Refinement (C3): Active parents should frequently co-activate with at least one active child.
- Conditional Correlation (C4): Parent activation magnitude should positively correlate with the maximum activation of its active children (addressing "soft" absorption).

Based on these requirements, the authors derive a unified, unsupervised evaluation protocol consisting of specific metrics (e.g., Monosemanticity Score, Hierarchical Abstractness Score, Activation Implication Score). They apply this protocol to evaluate six SAE approaches trained on visual data (CLIP ViT-L/14 and DINOv2 embeddings): a standard TopK SAE, Matryoshka SAEs (ActMSAE, ArchMSAE), EWG-SAE, Matching-Pursuit SAE (MP-SAE), and H-SAE.

Key Results
The evaluation reveals several critical findings regarding the state of hierarchical concept learning in SAEs:

Shallow Hierarchies: Most methods (TopK, EWG-SAE, ArchMSAE, H-SAE) produce predominantly two-level hierarchies. Only ActMSAE consistently generates deeper structures (more than two levels).
Semantic Structure vs. Activation Consistency: While feature spaces generally provide a basis for semantically structured graphs (high abstractness and sibling similarity scores), establishing consistent hierarchical activation behavior remains challenging.
The Absorption Problem: A central finding is the systematic compromise of hierarchy quality due to feature absorption.
- Hard Absorption: The suppression of parent activation when a child is active.
- Soft Absorption: A newly identified phenomenon where stronger child activations gradually suppress parent activations even when both are active, leading to poor correlation (low Conditional Correlation Score) and dominance scores.
Method-Specific Performance:
- Matryoshka approaches (ActMSAE): Show the best co-activation metrics and multi-level structures, though soft absorption persists.
- H-SAE (Explicit Hierarchy): Enforces consistent co-activation by design but fails to maintain semantic meaning; its features often lack interpretability, with children sometimes appearing more abstract than parents.
- Standard TopK SAE: Surprisingly forms meaningful semantic groupings but suffers significantly from inconsistent activation strengths between parents and children.
Encoder Independence: The challenges in learning hierarchical spaces appear driven by SAE training strategies rather than the underlying vision encoder (CLIP vs. DINOv2), provided the embeddings contain sufficient information.

Significance and Claims
The paper claims to establish the first general set of requirements and a corresponding quantitative evaluation protocol for unsupervised concept hierarchies. Its primary contribution is identifying that while SAEs can learn features capable of forming meaningful hierarchies, current approaches struggle to maintain the dynamic consistency required for these hierarchies to be practically useful.

The authors argue that the field must navigate a fundamental tension between sparsity and hierarchical activation consistency. They suggest that future approaches may need to rethink hard sparsity constraints, positing that a less sparse SAE with consistent hierarchical activation behavior could be more valuable than a sparser one with fragmented structure. Furthermore, they highlight that explicit structural constraints alone are insufficient without ensuring semantic coherence within the hierarchy.

The work concludes that while the proposed metrics and requirements can serve as training constraints for future extractors, open questions remain regarding the optimal method for extracting hierarchy graphs from latent spaces and the balance between explicit enforcement and implicit induction of structure.