Pattern Selectivity is Not Task-Causal Structure: A Cross-Architecture Mechanistic Study of Composed-Task Circuits in 1B-Class Language Models

In three linesMechanistic cross-architecture study on 3 1B-class models (Pythia, OLMo, OLMoE) testing whether circuit identification via pattern selectivity + causal ablation yields reproducible findings. Result: same task, same behavioral capability, different implementations across models. Five-category taxonomy (primary cause, secondary cause, correlate, interferer, null) with quantitative thresholds introduced.

Summary generated by Claude — human-verified