MIB: A Mechanistic Interpretability Benchmark
Fuente:
arXiv
Guardado en:
| Autores principales: | , , , , , , , , , , , , , , , , , , , , , , |
|---|---|
| Formato: | Preprint |
| Publicado: |
2025
|
| Materias: | |
| Acceso en línea: | |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| _version_ | 1866915333768478720 |
|---|---|
| author | Mueller, Aaron Geiger, Atticus Wiegreffe, Sarah Arad, Dana Arcuschin, Iván Belfki, Adam Chan, Yik Siu Fiotto-Kaufman, Jaden Haklay, Tal Hanna, Michael Huang, Jing Gupta, Rohan Nikankin, Yaniv Orgad, Hadas Prakash, Nikhil Reusch, Anja Sankaranarayanan, Aruna Shao, Shun Stolfo, Alessandro Tutek, Martin Zur, Amir Bau, David Belinkov, Yonatan |
| author_facet | Mueller, Aaron Geiger, Atticus Wiegreffe, Sarah Arad, Dana Arcuschin, Iván Belfki, Adam Chan, Yik Siu Fiotto-Kaufman, Jaden Haklay, Tal Hanna, Michael Huang, Jing Gupta, Rohan Nikankin, Yaniv Orgad, Hadas Prakash, Nikhil Reusch, Anja Sankaranarayanan, Aruna Shao, Shun Stolfo, Alessandro Tutek, Martin Zur, Amir Bau, David Belinkov, Yonatan |
| contents | How can we know whether new mechanistic interpretability methods achieve real improvements? In pursuit of lasting evaluation standards, we propose MIB, a Mechanistic Interpretability Benchmark, with two tracks spanning four tasks and five models. MIB favors methods that precisely and concisely recover relevant causal pathways or causal variables in neural language models. The circuit localization track compares methods that locate the model components - and connections between them - most important for performing a task (e.g., attribution patching or information flow routes). The causal variable localization track compares methods that featurize a hidden vector, e.g., sparse autoencoders (SAEs) or distributed alignment search (DAS), and align those features to a task-relevant causal variable. Using MIB, we find that attribution and mask optimization methods perform best on circuit localization. For causal variable localization, we find that the supervised DAS method performs best, while SAE features are not better than neurons, i.e., non-featurized hidden vectors. These findings illustrate that MIB enables meaningful comparisons, and increases our confidence that there has been real progress in the field. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2504_13151 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | MIB: A Mechanistic Interpretability Benchmark Mueller, Aaron Geiger, Atticus Wiegreffe, Sarah Arad, Dana Arcuschin, Iván Belfki, Adam Chan, Yik Siu Fiotto-Kaufman, Jaden Haklay, Tal Hanna, Michael Huang, Jing Gupta, Rohan Nikankin, Yaniv Orgad, Hadas Prakash, Nikhil Reusch, Anja Sankaranarayanan, Aruna Shao, Shun Stolfo, Alessandro Tutek, Martin Zur, Amir Bau, David Belinkov, Yonatan Machine Learning Artificial Intelligence Computation and Language How can we know whether new mechanistic interpretability methods achieve real improvements? In pursuit of lasting evaluation standards, we propose MIB, a Mechanistic Interpretability Benchmark, with two tracks spanning four tasks and five models. MIB favors methods that precisely and concisely recover relevant causal pathways or causal variables in neural language models. The circuit localization track compares methods that locate the model components - and connections between them - most important for performing a task (e.g., attribution patching or information flow routes). The causal variable localization track compares methods that featurize a hidden vector, e.g., sparse autoencoders (SAEs) or distributed alignment search (DAS), and align those features to a task-relevant causal variable. Using MIB, we find that attribution and mask optimization methods perform best on circuit localization. For causal variable localization, we find that the supervised DAS method performs best, while SAE features are not better than neurons, i.e., non-featurized hidden vectors. These findings illustrate that MIB enables meaningful comparisons, and increases our confidence that there has been real progress in the field. |
| title | MIB: A Mechanistic Interpretability Benchmark |
| topic | Machine Learning Artificial Intelligence Computation and Language |
| url | https://arxiv.org/abs/2504.13151 |