Saved in:
| Main Authors: | , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2502.17332 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866912244036534272 |
|---|---|
| author | Dooms, Thomas Wilhelm, Daniel |
| author_facet | Dooms, Thomas Wilhelm, Daniel |
| contents | Sparse auto-encoders (SAEs) have become a prevalent tool for interpreting language models' inner workings. However, it is unknown how tightly SAE features correspond to computationally important directions in the model. This work empirically shows that many RES-JB SAE features predominantly correspond to simple input statistics. We hypothesize this is caused by a large class imbalance in training data combined with a lack of complex error signals. To reduce this behavior, we propose a method that disentangles token reconstruction from feature reconstruction. This improvement is achieved by introducing a per-token bias, which provides an enhanced baseline for interesting reconstruction. As a result, significantly more interesting features and improved reconstruction in sparse regimes are learned. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2502_17332 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Tokenized SAEs: Disentangling SAE Reconstructions Dooms, Thomas Wilhelm, Daniel Machine Learning Sparse auto-encoders (SAEs) have become a prevalent tool for interpreting language models' inner workings. However, it is unknown how tightly SAE features correspond to computationally important directions in the model. This work empirically shows that many RES-JB SAE features predominantly correspond to simple input statistics. We hypothesize this is caused by a large class imbalance in training data combined with a lack of complex error signals. To reduce this behavior, we propose a method that disentangles token reconstruction from feature reconstruction. This improvement is achieved by introducing a per-token bias, which provides an enhanced baseline for interesting reconstruction. As a result, significantly more interesting features and improved reconstruction in sparse regimes are learned. |
| title | Tokenized SAEs: Disentangling SAE Reconstructions |
| topic | Machine Learning |
| url | https://arxiv.org/abs/2502.17332 |