What, Where, and How: Disentangling the Roles of Task, Language, and Model in Code Model Representations
Why different AI coding models build the same ideas in completely different ways
Two popular code-writing AI models agree on which programming concepts matter most, but they organize those concepts in their internal wiring in strikingly different ways. When tested on Python and Rust across both models, researchers found that what the models learn is universal, but how and where they learn it varies by model—suggesting that neural agreement is a property of ideas, not of computation itself.
Understanding how different AI models represent the same code concepts could improve model design, transfer learning, and interpretability. The finding that models share some neurons across languages but diverge in how they organize circuits reveals that current AI training doesn't converge on a single solution—which matters for predicting whether advances in one model will transfer to others, and for building more reliable code assistants.