Detailed Static and Dynamic Circuit Results across Transformer Models

This supplementary document provides detailed model-specific experimental results for static and dynamic circuits across Transformer models. The analyses cover 29 model configurations spanning natural language, programming languages, biomedical text, protein sequences, and molecular representations. Results are reported for static circuits, dynamic MLP circuits, dynamic attention–MLP circuits, and dynamic attention circuits under short- and long-context evaluation settings. For each model, the document presents target-promotion performance across circuit sparsity levels, including target probability, target logit, and rank-1 accuracy. The experiments compare top-ranked pathway selection with alternative selection or intervention conditions, including random baselines and pathway ablations. Additional counterfactual experiments evaluate dynamic MLP circuits across 11 Transformer architectures using input–target pairs that differ from the model's original predictions. These results provide model-level evidence complementing the aggregate analyses of sparse, target-conditioned computational pathways in Transformer models.