Polymorphism Is Rotation: Operational Mechanistic Interpretability from a Two-Layer Transformer to Pythia-70m
📰 ArXiv cs.AI
Learn how polymorphism in transformers can be understood as rotation, enabling operational mechanistic interpretability from small to large models like Pythia-70m
Action Steps
- Apply orthogonal Procrustes fit to remove polymorphism between model pairs
- Use a single batch of activations to transfer sparse-autoencoder feature dictionaries
- Analyze the effect of rotation on residual-stream bases in transformers
- Implement polymorphism removal using matrix multiplication
- Test the transferability of steering vectors between models
Who Needs to Know This
Researchers and engineers working with transformers and interpretability can benefit from understanding polymorphism as rotation to improve model understanding and transferability
Key Insight
💡 Polymorphism in transformers can be understood and removed as a uniform random rotation, enabling better model understanding and transferability
Share This
🤖 Polymorphism in transformers is just rotation! 🔄 Remove it with orthogonal Procrustes fit and improve model interpretability 📈
Key Takeaways
Learn how polymorphism in transformers can be understood as rotation, enabling operational mechanistic interpretability from small to large models like Pythia-70m
Full Article
Title: Polymorphism Is Rotation: Operational Mechanistic Interpretability from a Two-Layer Transformer to Pythia-70m
Abstract:
arXiv:2605.24577v1 Announce Type: cross Abstract: Independently trained transformers compute the same function in residual-stream bases that differ by a uniform random rotation on $\mathrm{SO}(d_{\mathrm{model}})$. We call this phenomenon polymorphism: same function, mutually unintelligible interior coordinates. One matrix multiplication per model pair removes it: an orthogonal Procrustes fit on a single batch of activations transfers sparse-autoencoder feature dictionaries and steering vectors
Abstract:
arXiv:2605.24577v1 Announce Type: cross Abstract: Independently trained transformers compute the same function in residual-stream bases that differ by a uniform random rotation on $\mathrm{SO}(d_{\mathrm{model}})$. We call this phenomenon polymorphism: same function, mutually unintelligible interior coordinates. One matrix multiplication per model pair removes it: an orthogonal Procrustes fit on a single batch of activations transfers sparse-autoencoder feature dictionaries and steering vectors
DeepCamp AI