Architecture, Not Scale: Circuit Localization in Large Language Models
For mechanistic interpretability researchers, it challenges the assumption that larger models are inherently harder to analyze, suggesting architectural choices can make large models more tractable.
The paper shows that attention architecture, not model scale, determines circuit concentration and stability in LLMs. Grouped query attention yields far more concentrated circuits than multi-head attention, and factual recall circuits undergo a discrete phase transition above a critical scale.
Mechanistic interpretability assumes that circuit analysis becomes harder as models scale. We challenge this assumption by showing that the attention architecture matters more than parameter count. Studying three circuit types across Pythia and Qwen2.5, we find that grouped query attention produces circuits that are far more concentrated and mechanistically stable than standard multi-head attention at comparable scales. The same concentration pattern holds across indirect object identification, induction heads, and factual recall. Within a single architecture family (Qwen2.5), factual recall circuits undergo a discrete phase transition above a critical scale, collapsing to a single bottleneck rather than degrading gradually. These findings suggest that some architectural choices make large models more tractable to study and that interpretability difficulty is not a fixed consequence of model size.