Tracr: Compiled Transformers as a Laboratory for Interpretability David Lindner

Feb-14-2026, 22:18:19 GMT–Neural Information Processing Systems 

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