Research question
Do sparse autoencoders resolve superposition in language models?
Key findings
- Dictionary features are more interpretable than neurons or PCA directions - Features enable precise causal interventions
Home/Research Paper Library/Sparse Autoencoders Find Highly Interpretable Features in Language Models
Sparse Autoencoders Find Highly Interpretable Features in Language Models
Independent, concurrent evidence that sparse autoencoders extract interpretable features from language models.
Contribute
Spot an error, or know a record we should add? Members can propose additions and submit corrections.