Code Circuits Are Universal in What, Not Where: Qwen and DeepSeek Agree on Which Concepts Get Circuits (rho ~0.65) but Place Them 11 Layers Apart

arXiv 2607.21491·medium signal

Piotr Wilam crossed Python and Rust with Qwen2.5-Coder-7B and DeepSeek-Coder-V1-6.7B, measuring a complete inventory of grammatical concepts (58 Python, 57 Rust) identically in all four cells. Which concepts earn dedicated circuitry is set by the task — the two models agree at Spearman rho = 0.638 for Python and 0.673 for Rust, both p < 1e-7 — but where those circuits sit is set by the model: Qwen processes in a late band around layers 17-19, DeepSeek at layers 6-7, for both languages. Rust constructs receive 2-3x more concept-specific circuitry than Python equivalents in both models, and Qwen binds nine Rust type-and-trait keywords into one neuron cluster (Jaccard 0.535 vs null 0.112, p < 0.001).