Sparse Dictionary Learning for Exhaustive Feature Discovery

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

Version 1 of 121 papers27 conceptsfrozen 2026-10-04

Neural networks are often described as "black boxes," but the field of mechanistic interpretability aims to change that by decomposing complex activations into human-understandable concepts. The current "gold standard" for this task is the Sparse Autoencoder (SAE). As detailed in a comprehensive survey by Dong Shu and colleagues, A Survey on Sparse Autoencoders, SAEs work by forcing a model to reconstruct its internal states using only a small number of active "features" from a large dictionary.

However, a growing body of evidence suggests that our current maps of these internal states are fundamentally incomplete. The central challenge in this research direction is not just finding some features, but ensuring we have found all of them—a property known as exhaustiveness.

The Exhaustiveness Problem

If an SAE identifies a feature for "The Golden Gate Bridge," we might feel we have successfully interpreted that part of the model. But as Adam Karvonen and his team argue in Measuring Progress in Dictionary Learning for Language Model Interpretability with Board Game Models, standard metrics often fail to tell us if we are missing a "long tail" of nuanced features.

This incompleteness has serious implications for AI safety. If we use SAEs to audit a model for bias, as explored by Zhengyang Shan in Measuring Mechanistic Independence, any bias-related mechanism residing in the "unmapped" portion of the activations will remain invisible and unmitigated.

Why Dictionaries Are Incomplete

The incompleteness of SAEs stems from three primary factors:

  1. Capacity Thresholds: Patrick Leask and colleagues found in Sparse Autoencoders Do Not Find Canonical Units of Analysis that smaller dictionaries don't just provide "blurry" versions of features; they omit specific concepts entirely. Dictionary size acts as a hard threshold for discovery.
  2. Stochasticity: Feature discovery is not an objective "truth" but a pragmatic choice. Nora Belrose and Gonçalo Paulo demonstrated in Sparse Autoencoders Trained on the Same Data Learn Different Features that SAEs trained with different random seeds on the same data might share as little as 30% of their features.
  3. Data Blindness: An SAE can only learn features present in its training corpus. Aly M. Kassem and colleagues show in REVIVING YOUR MNEME that if an audit dataset lacks niche concepts, the resulting dictionary will be blind to side effects affecting those concepts.

From Single Models to Ensembles

To combat this incompleteness, researchers are moving away from training single, massive SAEs toward ensembling. Chris Lin, Soham Gadgil, and Su-In Lee have shown that "more is better." In their work on Ensembling Sparse Autoencoders, they apply classic machine learning techniques like bagging and boosting to combine multiple SAEs. This approach significantly improves the reconstruction of neural activations and uncovers a far more diverse set of features than any single model could provide.

Discovery as the Primary Goal

Given these limitations, how should we use SAEs? Kenny Peng and colleagues suggest a shift in perspective in Use Sparse Autoencoders to Discover Unknown Concepts, Not to Act on Known Concepts. They argue that while SAEs are mediocre at "steering" concepts we already know (where supervised probes perform better), they are uniquely powerful for unsupervised discovery.

This power is being realized in diverse fields. For instance, Konstantin Donhauser and colleagues applied these methods to biological foundation models in Towards scientific discovery with dictionary learning, extracting interpretable biological concepts like cell types from raw microscopy images.

Turning Gaps into Signals

Perhaps the most intuitive leap in this field is treating the "missing" information not as a failure, but as a signal. Yavuz Faruk Bakman and colleagues propose in Uncertainty as Feature Gaps that we can quantify a model's uncertainty by looking at the "gaps" in its internal feature space. Similarly, Maty Bohacek and colleagues use these gaps to identify Competency Gaps in LLMs, mapping exactly where a model's internal conceptual map fails to cover real-world topics like legal compliance.

By acknowledging that our dictionaries are currently incomplete, researchers are developing more robust ensembling methods and more honest ways to measure what an AI truly "knows."

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_sparse_dictionary_learning_for_exhaustiv_e19f2e39,
  title = {Sparse Dictionary Learning for Exhaustive Feature Discovery},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/sparse-dictionary-learning-for-exhaustive-feature-discovery/rdr_e19f2e3975f44821a2aaaa883413d298},
  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_sparse_dictionary_learning_for_exhaustiv_e19f2e39_v1,
  title = {Sparse Dictionary Learning for Exhaustive Feature Discovery},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/sparse-dictionary-learning-for-exhaustive-feature-discovery/rdr_e19f2e3975f44821a2aaaa883413d298/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}