IC-808Sparse autoencoder features in Pythia-70m's residual stream are more interpretable than PCA, ICA, random, and default-basis directions, with the advantage declining from early to late layers

Robert Huben, Hoagy Cunningham, Logan Riggs Smith, Aidan Ewart, Lee Sharkey

SourceSparse Autoencoders Find Highly Interpretable Features in Language Models

The authors train sparse autoencoders on Pythia-70m's residual stream activations and measure the interpretability of the learned directions using the autointerpretability protocol (GPT-4 generates a description, GPT-3.5 simulates activations, correlation is the score). Across all layers, the sparse dictionary features score substantially higher than PCA, ICA, random directions, and the default neuron basis. However, the gap narrows with depth: by layer 4 the sparse features are comparable to ICA, and in the final layer the improvement over baselines is minimal. The authors note this may reflect the fixed hyperparameters (alpha=0.00086, r=2) producing more active features in later layers, or that later-layer features are inherently harder to interpret automatically.

Evidence
correlational
Caveat
The autointerpretability metric relies on GPT-4/GPT-3.5 as judges, and the authors note that current LLMs are limited in the kinds of patterns they can find (e.g., patterns centred on next/previous tokens). The fixed hyperparameters may not be optimal for later layers.
Model
Pythia Pythia-70m
Concepts
Linear representation, Depth-dependent structure
Datasets
The Pile [source], OpenWebText / OpenWebText-10k [eval]
Methods
Sparse autoencoder / Sparse autoencoders / K-sparse autoencoder / Topk sparse autoencoder / Scaling and Evaluating Sparse Autoencoders / Cunningham et al. 2023 (sparse autoencoders) [primary], Autointerpretability [eval], Principal component analysis [compared-to]
Related work
Bills et al. 2023 (Language Models Can Explain Neurons in Language Models) [builds-on], Sharkey et al. 2023 (Taking Features Out of Superposition with Sparse Autoencoders) [builds-on], Elhage et al. 2022b (Toy Models of Superposition) [context]
Related findings
IC-809, IC-810
Extraction
automatic-extraction