Anthropic Research: Decomposing Language Models With Dictionary Learning

Anthropic has introduced a method for decomposing language models using dictionary learning to identify "features"—linear combinations of neuron activations—that represent specific, interpretable concepts. This approach allows researchers to understand the internal representations of a transformer model by breaking down high-dimensional activations into monosemantic units of analysis.

Dictionary Learning and the Concept of Monosemanticity

Dictionary learning is used to decompose a layer of neurons into a larger number of features. In the provided research, Anthropic decomposed a layer with 512 neurons into more than 4,000 features. This process reveals that a single neuron may not be the most effective unit of analysis for an AI model's internal logic, as many model properties are invisible when observing individual neurons in isolation.

Interpretable Features in Transformer Models

By identifying these features, researchers can move from polysemantic neurons (neurons that fire for multiple unrelated concepts) to monosemantic features (units that represent a single, clear concept). The identified features in their small transformer model correspond to a specific range of diverse concepts, including:

• DNA sequences
• Legal language
• HTTP requests
• Hebrew text
• Nutrition statements

Implications for Neural Network Interpretability

This research builds on existing efforts to interpret high-dimensional systems across neuroscience, machine learning, and statistics. By providing a machinery to find these units of analysis, Anthropic aims to create a path toward breaking down complex neural networks into parts that are human-understandable, increasing the transparency of how frontier models process information.