Tracr: Compiled Transformers as a Laboratory for Interpretability David Lindner

Oct-8-2025, 22:39:00 GMT–Neural Information Processing Systems 

We show how to "compile" human-readable programs into standard decoder-only transformer models.