Tag
This paper investigates neural circuits in a sparse 8-layer Python transformer, finding dedicated circuitry for 106 programming concepts and decomposing them into concept-specific and token-driven components, with implications for understanding structural encoding in code models.