Towards Monosemanticity: Decomposing Language Models With Dictionary Learning

Paper · Source
Mechanistic Interpretability

Mechanistic interpretability seeks to understand neural networks by breaking them into components that are more easily understood than the whole. By understanding the function of each component, and how they interact, we hope to be able to reason about the behavior of the entire network. The first step in that program is to identify the correct components to analyze.

Unfortunately, the most natural computational unit of the neural network – the neuron itself – turns out not to be a natural unit for human understanding. This is because many neurons are polysemantic: they respond to mixtures of seemingly unrelated inputs. In the vision model Inception v1, a single neuron responds to faces of cats and fronts of cars [1] . In a small language model we discuss in this paper, a single neuron responds to a mixture of academic citations, English dialogue, HTTP requests, and Korean text. Polysemanticity makes it difficult to reason about the behavior of the network in terms of the activity of individual neurons.

One potential cause of polysemanticity is superposition [2, 3, 4, 5] , a hypothesized phenomenon where a neural network represents more independent "features" of the data than it has neurons by assigning each feature its own linear combination of neurons. If we view each feature as a vector over the neurons, then the set of features form an overcomplete linear basis for the activations of the network neurons. In our previous paper on Toy Models of Superposition [5] , we showed that superposition can arise naturally during the course of neural network training if the set of features useful to a model are sparse in the training data. As in compressed sensing, sparsity allows a model to disambiguate which combination of features produced any given activation vector. 1

Synthesis notes that discuss concepts related to this paper 7

Notes in this collection most semantically related to the paper, ranked by cosine similarity.

Lines of inquiry this paper opens 24

Research framings built by reading the notes related to this paper — the questions it feeds into.

What determines how neural networks organize their internal knowledge representations? How can parameter sparsity improve interpretability while maintaining model capability? Can scaling alone produce compositional generalization in neural networks? How do different post-training structures affect learning outcomes? What distinguishes genuine understanding from correct linguistic performance? Do language models learn underlying linguistic structure or merely surface patterns? How do pretraining data composition and task diversity shape reasoning? How do transformer circuits and attention mechanisms enable or limit reasoning? Why do language models systematically fail at causal reasoning? Does architectural recurrence enable computational capabilities that fixed-depth models cannot achieve? How do mechanistic features and representations drive model behavior?