Scaling Neural Interpretability with Automated Feature Labeling

This direction is merged into another direction: interpretability scalability. Its last version stays readable and citable here.

Version 1 of 131 papers32 conceptsfrozen 2026-10-04

The central challenge of mechanistic interpretability is a matter of scale. While we can manually inspect a handful of neurons to understand their function, modern Large Language Models (LLMs) and Vision Transformers (ViTs) contain millions of features. As Liu Leqi and colleagues argue in Diagnostic Curves for Black Box Models, manual inspection becomes "computationally and cognitively infeasible" as models grow. To bridge this gap, researchers are developing systems that use models to explain other models, automating the discovery and labeling of neural concepts.

The Bottleneck of Polysemanticity

Before a model can be explained, its internal activations must be "de-noised." Individual neurons are often polysemantic, meaning a single neuron might fire for multiple unrelated concepts (e.g., both "cats" and "quantum physics"). This makes direct labeling impossible.

To solve this, researchers like Samuel Stevens and Matthew Bozoukov employ Sparse Autoencoders (SAEs). In Interpretable and Testable Vision Features via Sparse Autoencoders, the authors use SAEs to disentangle these messy activations into a larger set of "sparse" features, each representing a single, human-understandable concept. Bozoukov demonstrated in Uncovering Branch-specialization in InceptionV1 using k sparse autoencoders that this disentanglement reveals how specific model branches specialize in distinct categories, such as animal species or textures.

Models Explaining Models

Once features are disentangled, the next hurdle is assigning them names. Historically, this required a human to look at "top-activating examples" and guess the pattern—a process Sandeep Konam and Manuela Veloso sought to automate with APPLE.

The modern approach uses frontier LLMs as "automated labellers." In Automatically Interpreting Millions of Features in Large Language Models, Gonçalo Paulo and the EleutherAI team describe a pipeline where a "teacher" model (like GPT-4) examines the data that triggers a specific feature and generates a natural language explanation. This allows researchers to move from labeling dozens of neurons to millions of features.

This isn't limited to text. Chenxu Zhao and Wei Qian show in Automated Natural Language Explanation of Deep Visual Neurons with Large Models that Vision-Language Models (VLMs) can provide precise descriptions of visual neurons, eliminating the need for human-labeled datasets in computer vision auditing.

From Features to Circuits

Understanding individual features is only the first step; the ultimate goal is understanding the circuits—the logic gates formed by interacting features.

Kaichen Zhang and Ziwei Liu demonstrate in Large Multi-modal Models Can Interpret Features in Large Multi-modal Models that once we have automated labels, we can "steer" a model's reasoning by amplifying or suppressing specific features to correct errors like hallucinations.

Similarly, frameworks like NeurFlow, developed by Tue Minh Cao and My T. Thai, move beyond individual units to identify "core concept neurons" and cluster them into functional groups. In NeurFlow: Interpreting Neural Networks through Neuron Groups and Functional Interactions, they show how these groups interact across layers to build complex concepts from simple ones.

To make these interactions readable, Wenlong Yu and Qilong Wang introduced Chain-of-Explanation (CoE), which converts these internal circuits into step-by-step linguistic narratives. Instead of a heat map, the researcher receives a narrative: "The model first detected a circular shape, then identified it as a wheel, and finally classified the object as a bicycle."

Why Automation Matters

The shift toward automated interpretability isn't just about speed; it's about systematicity. As Angie Boggust and Hendrik Strobelt highlight with their Shared Interest framework, manual inspection is prone to selection bias. By automating the measurement of alignment between model behavior and human concepts, we can identify "shortcuts" or biases—like a sentiment classifier that only looks for the word "check"—that would be missed in a small-scale manual audit.

By treating interpretability as a scalable engineering problem rather than a manual art, this research direction aims to make the "black box" of AI transparent enough to audit, debug, and safely deploy.

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_scaling_neural_interpretability_with_aut_41836603,
  title = {Scaling Neural Interpretability with Automated Feature Labeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/scaling-neural-interpretability-with-automated-feature-labeling/rdr_418366030ec143c9800757b7caf904c1},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_scaling_neural_interpretability_with_aut_41836603_v1,
  title = {Scaling Neural Interpretability with Automated Feature Labeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/scaling-neural-interpretability-with-automated-feature-labeling/rdr_418366030ec143c9800757b7caf904c1/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}