Euan Ong

Successor Heads: Recurring, Interpretable Attention Heads In The Wild

Compact Proofs of Model Performance via Mechanistic Interpretability

Image Hijacks: Adversarial Images can Control Generative Models at Runtime