Saved in:
| Main Authors: | , , , , , , , , , , , , , , , , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2605.29358 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866910269023715328 |
|---|---|
| author | Templeton, Adly Conerly, Tom Marcus, Jonathan Lindsey, Jack Bricken, Trenton Chen, Brian Pearce, Adam Citro, Craig Ameisen, Emmanuel Jones, Andy Cunningham, Hoagy Turner, Nicholas L McDougall, Callum MacDiarmid, Monte Tamkin, Alex Durmus, Esin Hume, Tristan Mosconi, Francesco Freeman, C. Daniel Sumers, Theodore R. Rees, Edward Batson, Joshua Jermyn, Adam Carter, Shan Olah, Chris Henighan, Tom |
| author_facet | Templeton, Adly Conerly, Tom Marcus, Jonathan Lindsey, Jack Bricken, Trenton Chen, Brian Pearce, Adam Citro, Craig Ameisen, Emmanuel Jones, Andy Cunningham, Hoagy Turner, Nicholas L McDougall, Callum MacDiarmid, Monte Tamkin, Alex Durmus, Esin Hume, Tristan Mosconi, Francesco Freeman, C. Daniel Sumers, Theodore R. Rees, Edward Batson, Joshua Jermyn, Adam Carter, Shan Olah, Chris Henighan, Tom |
| contents | We demonstrate that sparse autoencoders can extract interpretable features from Claude 3 Sonnet, a production-scale language model, addressing the open question of whether dictionary learning methods scale beyond small transformers. We trained sparse autoencoders with up to 34 million features on the model's middle layer residual stream, using scaling laws to guide hyperparameter selection. The resulting features are multilingual and multimodal (generalizing to images despite text-only training), respond to both concrete instances and abstract discussions of concepts, and can be used to steer model behavior in ways consistent with their interpretations. We find features corresponding to famous entities and locations, as well as more abstract concepts like sarcasm or errors in code. We also identify features relevant to ways in which language models might cause harm--including features representing deception, power-seeking, sycophancy, and bias--and show that these causally influence model outputs when manipulated. Additionally, we conduct analyses of feature interpretability, geometry, and computational function. However, significant limitations remain: our suite of features is incomplete, and we lack rigorous methods for evaluating whether our features faithfully capture model computations. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_29358 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet Templeton, Adly Conerly, Tom Marcus, Jonathan Lindsey, Jack Bricken, Trenton Chen, Brian Pearce, Adam Citro, Craig Ameisen, Emmanuel Jones, Andy Cunningham, Hoagy Turner, Nicholas L McDougall, Callum MacDiarmid, Monte Tamkin, Alex Durmus, Esin Hume, Tristan Mosconi, Francesco Freeman, C. Daniel Sumers, Theodore R. Rees, Edward Batson, Joshua Jermyn, Adam Carter, Shan Olah, Chris Henighan, Tom Artificial Intelligence We demonstrate that sparse autoencoders can extract interpretable features from Claude 3 Sonnet, a production-scale language model, addressing the open question of whether dictionary learning methods scale beyond small transformers. We trained sparse autoencoders with up to 34 million features on the model's middle layer residual stream, using scaling laws to guide hyperparameter selection. The resulting features are multilingual and multimodal (generalizing to images despite text-only training), respond to both concrete instances and abstract discussions of concepts, and can be used to steer model behavior in ways consistent with their interpretations. We find features corresponding to famous entities and locations, as well as more abstract concepts like sarcasm or errors in code. We also identify features relevant to ways in which language models might cause harm--including features representing deception, power-seeking, sycophancy, and bias--and show that these causally influence model outputs when manipulated. Additionally, we conduct analyses of feature interpretability, geometry, and computational function. However, significant limitations remain: our suite of features is incomplete, and we lack rigorous methods for evaluating whether our features faithfully capture model computations. |
| title | Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet |
| topic | Artificial Intelligence |
| url | https://arxiv.org/abs/2605.29358 |