Towards Monosemanticity: Decomposing Language Models With Dictionary Learning

"Towards Monosemanticity: Decomposing Language Models With Dictionary Learning" (Trenton Bricken, Adly Templeton, Chris Olah, and colleagues at Anthropic, 2023) is the paper that showed sparse autoencoders could pull readable features out of a language model, and it's the starting point for much of the interpretability work described in the Interpretability lesson.

What problem it solved

The natural way to understand a network is to ask what each neuron does. That mostly fails for language models, because neurons are polysemantic: a single neuron may respond to unrelated things, such as academic citations, English dialogue, and HTTP requests, all at once. The leading explanation is superposition: the model needs to represent far more concepts than it has neurons, so it packs many features into overlapping combinations of them. Reading individual neurons then tells you almost nothing, and without a readable unit of analysis, claims about what a model "knows" or intends can't be checked.

The key idea

If features are packed into overlapping directions, find the directions instead of the neurons. The paper trained a sparse autoencoder on the activations of the MLP layer in a small, one-layer transformer: a separate network that reconstructs those activations through a much wider hidden layer, with a penalty that makes most of its units zero for any given input. The sparsity forces each input to be explained by only a few active units, and the width gives the model room to give each feature its own unit, unscrambling the superposition.

The authors then did the unglamorous part that makes the result credible: they inspected the features by hand and with automated tests. The recovered features turned out to be far more interpretable than the neurons they came from, with clean, specific meanings such as text in Arabic script, DNA sequences, base64 strings, and particular kinds of legal language, and changing a feature's activation changed the model's output in the expected way.

Why it mattered

It supplied a working tool for the central obstacle in interpretability, and it did so on real model activations rather than a toy. A follow-up, Scaling Monosemanticity (2024), applied the same recipe to a production-scale model and found features tied to concepts such as specific places, code bugs, and deception. The honest caveats are the ones the lesson stresses: this paper studied a one-layer model, the features it finds are a useful decomposition rather than a proven ground truth, and a feature that looks clean is evidence about what the model represents, not a guarantee about how it will behave. It is a method for looking, not a finished account of what's inside.

Authors: Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nick Turner, Cem Anil, Carson Denison, Amanda Askell, Robert Lasenby, Yifan Wu, Shauna Kravec, Nicholas Schiefer, Tim Maxwell, Nicholas Joseph, Zac Hatfield-Dodds, Alex Tamkin, Karina Nguyen, Brayden McLean, Josiah E. Burke, Tristan Hume, Shan Carter, Tom Henighan, Chris Olah (Anthropic)

Read the paper — transformer-circuits.pub

Learn more: Sparse Autoencoder (SAE) · Interpretability · AI Safety & Alignment