sae features universal rotation lesswrong
A LessWrong post examines whether sparse autoencoder features are consistent across models aside from an unknown rotation factor.
Topic cluster
research method
3 sources grouped by AFBytes in Ai
AFBytes briefing
Advances in AI interpretability can influence how developers and regulators assess model reliability and safety.
What to watch next
- Watch for follow-up papers or open-source releases that quantify rotation effects in production models.
Ablating Induction Heads Increases Local Repetition in Models
An independent project tests how removing specific attention heads alters repetition behavior in language models.
Tensor similarity method for neural network comparison
Researchers introduced a weight-based method to quantify functional similarity between neural networks across inputs.