TraCR-Supported Mechanistic Interpretability We compare a Transformer compiled from a RASP algorithm using TRACR to a Transformer trained on the same task with AdamW using the TransformerLens library. We find that (1) compiled Transformers...