Deep neural networks are notoriously opaque, representing information in high-dimensional "latent spaces" that defy direct human inspection. While we can observe that a specific neuron fires, understanding what it represents—whether a grammatical rule, a visual style, or a biological property—has traditionally required painstaking manual effort.
A rapidly converging research direction is automating this process. By combining decomposition techniques like Sparse Autoencoders (SAEs) with the descriptive power of Large Language Models (LLMs), researchers are building pipelines that can "crack open" these black boxes, assigning human-readable labels to millions of internal features at scale.
The Decomposition Bottleneck
The fundamental challenge in mechanistic interpretability is polysemanticity: a single neuron often responds to multiple, unrelated concepts. To solve this, researchers use methods like Sparse Autoencoders to project dense activations into a much higher-dimensional, sparse space where each "feature" is more likely to represent a single, monosemantic concept.
However, even a perfectly sparse model might have millions of features. Gonçalo Paulo and colleagues at EleutherAI note in Automatically Interpreting Millions of Features in Large Language Models that manual annotation is impossible at this scale. The solution is to treat interpretability as a generative task: using a secondary "explainer" model to synthesize the commonalities among the inputs that trigger a specific feature.
The Automated Labeling Pipeline
The emerging standard for automated interpretation follows a three-step loop:
- Feature Extraction: A model’s internal states are decomposed into sparse components using SAEs, Independent Component Analysis (ICA), or matrix factorization.
- Activation Sampling: The system identifies the "top-activating" examples—the specific text snippets, images, or data points that cause a particular latent feature to fire most strongly.
- Semantic Synthesis: A high-capacity LLM analyzes these examples and generates a concise natural language description of the underlying concept.
This approach is remarkably versatile. In the realm of computer vision, Rameen Abdal and Niloy Mitra developed CLIP2StyleGAN, which uses the multimodal space of CLIP to automatically name editing directions in StyleGAN, turning abstract vectors into commands like "aging" or "wrecked car." Similarly, Shaonan Wang uses LLMs to interpret "semantic axes" in word embeddings, mapping them to human-interpretable descriptors like "static vs. dynamic" Discovering Semantic Subdimensions through Disentangled Conceptual Representations.
Beyond Text: Specialized Domains
Automated labeling is particularly powerful in technical fields where human intuition is limited. Elana P. Simon applied this to protein language models in InterPLM, using SAEs to discover thousands of biological features—like specific binding sites—that were then automatically labeled using biological metadata.
In healthcare, Hiba Ahsan and Byron C Wallace demonstrated that automated pipelines can reveal hidden racial biases in medical LLMs that are not apparent in the model's final output Can SAEs reveal and mitigate racial biases of LLMs in healthcare?. By generating domain-specific descriptions for internal features, they could identify where a model had internalized harmful stereotypes, such as linking specific identities to incarceration or drug use.
From Discovery to Utility
The ultimate goal of automated labeling isn't just to describe models, but to control them. Once a feature is labeled, it becomes a "handle" for steering. Nicholas Jiang and Neel Nanda propose using these labeled features as a permanent, reusable dictionary for data analysis, allowing researchers to query massive datasets using human-readable concepts rather than raw vectors Interpretable Embeddings with Sparse Autoencoders: A Data Analysis Toolkit.
Similarly, Rajiv Movva and Emma Pierson introduced WIMHF What's In My Human Feedback?, which uses this pipeline to audit the preference data used to train LLMs. By automatically labeling features like "uses emojis" or "refuses request," they can identify misaligned data and improve model safety.
The Fidelity Challenge
While automated labeling scales interpretability, it introduces a new question: how do we know the labels are accurate? Benyou Wang and colleagues caution that "higher interpretability" (as measured by an LLM's ability to label a feature) does not always correlate with the feature's actual utility in steering the model Does higher interpretability imply better utility?.
To address this, researchers like Jiayin Wang are developing "closed-loop" verification systems, where the explainer model must use its own generated labels to predict whether the feature will activate on new, unseen inputs Interpret the Internal States of Recommendation Model with Sparse Autoencoder. This ensures that the natural language description isn't just a plausible guess, but a faithful map of the model's internal logic.