This repo contains the code for the paper Sparse but Wrong: Incorrect L0 Leads to Incorrect Features in Sparse Autoencoders.

This project uses uv for package management. You can install dependencies with:

uv sync

Or, you can just use pip and run:

pip install -e .

The toy model experiments are all in the notebooks directory, with supporting files in the toy_model directory. We also provide a demo notebook for how to reproduce our Gemma-2-2b experiments in notebooks/train_and_eval_llm_sae.ipynb.

We extend the SAELens BatchTopKSAE class in enhanced_batch_topk_sae.py to keep the decoder normalized and support our experiments where we change the L0 of the SAE during training.

We use pytest for testing. You can run the tests with:

uv run pytest

We also use ruff for linting and pyright for type checking. You can run the linting and type checking with:

uv run ruff check .
uv run pyright

@article{chanin2025sparse,
     title={Sparse but Wrong: Incorrect L0 Leads to Incorrect Features in Sparse Autoencoders},
     author={David Chanin and Adrià Garriga-Alonso},
     year={2025},
     journal={arXiv preprint arXiv:2508.16560}
}