SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2025
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866910985357361152 |
|---|---|
| author | Karvonen, Adam Rager, Can Lin, Johnny Tigges, Curt Bloom, Joseph Chanin, David Lau, Yeu-Tong Farrell, Eoin McDougall, Callum Ayonrinde, Kola Till, Demian Wearden, Matthew Conmy, Arthur Marks, Samuel Nanda, Neel |
| author_facet | Karvonen, Adam Rager, Can Lin, Johnny Tigges, Curt Bloom, Joseph Chanin, David Lau, Yeu-Tong Farrell, Eoin McDougall, Callum Ayonrinde, Kola Till, Demian Wearden, Matthew Conmy, Arthur Marks, Samuel Nanda, Neel |
| contents | Sparse autoencoders (SAEs) are a popular technique for interpreting language model activations, and there is extensive recent work on improving SAE effectiveness. However, most prior work evaluates progress using unsupervised proxy metrics with unclear practical relevance. We introduce SAEBench, a comprehensive evaluation suite that measures SAE performance across eight diverse metrics, spanning interpretability, feature disentanglement and practical applications like unlearning. To enable systematic comparison, we open-source a suite of over 200 SAEs across eight recently proposed SAE architectures and training algorithms. Our evaluation reveals that gains on proxy metrics do not reliably translate to better practical performance. For instance, while Matryoshka SAEs slightly underperform on existing proxy metrics, they substantially outperform other architectures on feature disentanglement metrics; moreover, this advantage grows with SAE scale. By providing a standardized framework for measuring progress in SAE development, SAEBench enables researchers to study scaling trends and make nuanced comparisons between different SAE architectures and training methodologies. Our interactive interface enables researchers to flexibly visualize relationships between metrics across hundreds of open-source SAEs at: www.neuronpedia.org/sae-bench |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2503_09532 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability Karvonen, Adam Rager, Can Lin, Johnny Tigges, Curt Bloom, Joseph Chanin, David Lau, Yeu-Tong Farrell, Eoin McDougall, Callum Ayonrinde, Kola Till, Demian Wearden, Matthew Conmy, Arthur Marks, Samuel Nanda, Neel Machine Learning Computation and Language Sparse autoencoders (SAEs) are a popular technique for interpreting language model activations, and there is extensive recent work on improving SAE effectiveness. However, most prior work evaluates progress using unsupervised proxy metrics with unclear practical relevance. We introduce SAEBench, a comprehensive evaluation suite that measures SAE performance across eight diverse metrics, spanning interpretability, feature disentanglement and practical applications like unlearning. To enable systematic comparison, we open-source a suite of over 200 SAEs across eight recently proposed SAE architectures and training algorithms. Our evaluation reveals that gains on proxy metrics do not reliably translate to better practical performance. For instance, while Matryoshka SAEs slightly underperform on existing proxy metrics, they substantially outperform other architectures on feature disentanglement metrics; moreover, this advantage grows with SAE scale. By providing a standardized framework for measuring progress in SAE development, SAEBench enables researchers to study scaling trends and make nuanced comparisons between different SAE architectures and training methodologies. Our interactive interface enables researchers to flexibly visualize relationships between metrics across hundreds of open-source SAEs at: www.neuronpedia.org/sae-bench |
| title | SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability |
| topic | Machine Learning Computation and Language |
| url | https://arxiv.org/abs/2503.09532 |