Mechanistic interpretability project from the Mechanistic Interpretability Hackathon that studies properties of soft-tuning or prompt-tuning and empirically shows that task tokens can form a convex set.
Endorsements made here support MentaLeap.
Mechanistic interpretability project from the Mechanistic Interpretability Hackathon that studies properties of soft-tuning or prompt-tuning and empirically shows that task tokens can form a convex set.
Endorsements made here support MentaLeap.
People
Updated 05/18/26 · By grantmaking.aiTeam member
Discussion
Sign in to comment
No comments yet. Be the first to share your thoughts.