Logit Prisms: Decomposing Transformer Outputs for Mechanistic Interpretability June 17, 2024 by Comments