Original sparse autoencoder checkpoints and model resources used by SAEBase.
AI & ML interests
Building open foundations for understanding and evaluating LLMs.
Recent Activity
View all activity
Papers
Behavioral Compositionality of Sparse Autoencoder Representations
Representation Redistribution in Sparse Autoencoders
Organization Card
Open research in mechanistic interpretability
From features
to testable
explanations.
Open research methods and infrastructure for studying neural representations and model behavior.
SAEBase
One place to study what features do.
Explore feature interventions and reproducible results across 48 SAE sites and a 500-prompt research cohort.
Open the research lab1Capture
2Intervene
3Evaluate
18Model–SAE configurations
48Evaluated layer sites
Open science
Research should be easy to examine. And easier to build on.
models 0
None public yet