Hierarchical Feature Absorption in Sparse Autoencoders

This direction is merged into another direction: sparse feature decomposition. Its last version stays readable and citable here.

Version 1 of 116 papers29 conceptsfrozen 2026-10-04

Mechanistic interpretability aims to decompose the "black box" of Large Language Models (LLMs) into human-understandable components. The primary tool for this is the Sparse Autoencoder (SAE), which attempts to find a set of sparse, linear directions—features—that explain model activations.

However, recent research suggests that our current SAEs are biased by their own optimization objectives. Because LLMs represent knowledge in complex hierarchies, the mathematical pressure to be "sparse" often forces these models to merge distinct concepts into single, messy latents. This phenomenon, known as feature absorption, suggests that we are often looking at the internal state of AI through a lens that is out of focus.

The Mathematical Incentive to Cheat

The goal of an SAE is to reconstruct a model's activations using as few active features as possible. This is typically enforced using an $L_1$ or $L_0$ penalty. While this penalty encourages monosemanticity (one feature = one concept), it also creates a perverse incentive.

As Patrick Leask and colleagues argue in Sparse Autoencoders Do Not Find Canonical Units of Analysis, sparsity penalties incentivize the learning of composite features over atomic ones. Imagine a model needs to represent the concept of a "Golden Retriever." In a truly atomic system, it might activate two features: "Dog" and "Golden." But under a sparsity penalty, activating two features is twice as "expensive" as activating one. The SAE can minimize its loss by creating a single "Golden Retriever" feature that absorbs the semantic direction of "Dog."

This isn't just a theoretical quirk; it is a fundamental tension in how we measure interpretability. Representing a complex idea with one latent is mathematically prioritized over using multiple simpler, more fundamental building blocks.

Feature Absorption and Hedging

When an SAE is forced to operate under a strict sparsity budget, it engages in what David Chanin and Adrià Garriga-Alonso call feature hedging. In their work Sparse but Wrong: Incorrect L0 Leads to Incorrect Features in Sparse Autoencoders, they demonstrate that there is a "correct" level of sparsity ($L_0$) for any given layer. If the budget is set too low, the SAE "cheats" by merging correlated features to improve reconstruction scores, resulting in polysemantic latents that lack clear meaning.

This leads directly to feature absorption. In A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoders, David Chanin and Tomáš Dulka show that when features form a hierarchy, the child latent (the specific concept) often swallows the direction of the parent latent (the general concept). For example, a feature for "the word 'short'" might absorb the general feature for "words starting with 's'." Consequently, the general "parent" feature may appear to "fail" or disappear entirely because its signal has been distributed across its many "children."

The Geometric Reality of Hierarchy

Why does this happen so easily? The answer lies in how LLMs geometrically organize information. Kiho Park and Victor Veitch demonstrate in The Geometry of Categorical and Hierarchical Concepts in Large Language Models that LLMs represent categorical concepts as geometric polytopes.

Crucially, they find that hierarchical relationships are often encoded through orthogonality. A general concept (like "Animal") provides a base direction, while specific sub-concepts (like "Cat" or "Dog") exist in orthogonal subspaces. Standard SAEs assume a "flat" world where features are independent. When they encounter this hierarchical, orthogonal geometry, they struggle to distinguish between a shared parent component and the unique features of a child, leading to the "blurred" or absorbed representations noted by Xiaoqing Sun and Neel Nanda in their work on interpretable embeddings.

Toward Hierarchy-Aware Interpretability

Recognizing that the "flat" assumption is the bottleneck, researchers are proposing new architectures that respect the model's internal hierarchy.

One promising direction is MP-SAE, introduced by Valérie Costa and colleagues in From Flat to Hierarchical: Extracting Sparse Representations with Matching Pursuit. Instead of trying to find all features simultaneously, MP-SAE uses the Matching Pursuit algorithm to sequentially extract features from residuals. By "peeling off" features one by one, the model can successfully uncover hierarchical and nonlinear concepts that standard, flat SAEs miss.

Similarly, Marco Fumero and Stefano Soatto suggest that true disentanglement requires balancing sparse sufficiency (using few features) with minimality (sharing features across tasks) 2023. Without a mechanism to encourage feature sharing, models will continue to create redundant or absorbed representations.

Ultimately, understanding the "periodic table" of AI requires us to move past simple sparsity. We must design tools that are as geometrically sophisticated as the models they are meant to explain, ensuring that the "microscope" doesn't accidentally merge the very atoms it is trying to see.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_hierarchical_feature_absorption_in_spars_3629dbd0,
  title = {Hierarchical Feature Absorption in Sparse Autoencoders},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hierarchical-feature-absorption-in-sparse-autoencoders/rdr_3629dbd00207487387b5836b98af16fb},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_hierarchical_feature_absorption_in_spars_3629dbd0_v1,
  title = {Hierarchical Feature Absorption in Sparse Autoencoders},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hierarchical-feature-absorption-in-sparse-autoencoders/rdr_3629dbd00207487387b5836b98af16fb/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}