Building open foundations for understanding and evaluating LLMs.
🌐 Official Website · 🤗 Hugging Face
Research needs standard infrastructure.
OpenSAE is an open research non-profit organization building shared infrastructure for mechanistic interpretability. We develop reusable tools for understanding models, testing interventions, and evaluating explanations, so researchers can spend less time rebuilding software and more time testing ideas.
Make the tools common. Keep the insights unique.
Model inspection, feature editing, and evaluation should be accessible, reproducible, and reusable across research projects.
From a learned feature to a controlled experiment.
SAEBase is OpenSAE's main software platform for exploring, intervening in, and evaluating LLM representations. It offers an interactive workspace, Python tools, and research workflows.
Workspace ↗ · Python API ↗ · Documentation ↗ · Source ↗
Explanations need evidence.
We study what learned features represent, how they interact, how they change across contexts, and how they influence model behavior.
- 🧩 Behavioral Compositionality of Sparse Autoencoder Representations — When do combinations of features change model behavior?
- 🔀 Representation Redistribution in Sparse Autoencoders — How does feature activity change across contexts?
- 🧭 Sparse Autoencoders Do Not Guarantee Local Features — When is a learned feature genuinely local?
We welcome open-source contributions and research collaborations. Explore our repositories, open an issue, or get in touch at hma232@wisc.edu.
