Code models have a well-known default: when a prompt does not specify a programming language, they often generate Python. A new arXiv preprint (2610.20683) asks a mechanistic question about that behaviour—whether it is driven by a small, identifiable set of internal components rather than spread across the network.

The authors report running a single pipeline to search for such components. Their key finding, as stated in the abstract, is that whether the behaviour looks localised depends on the 'budget' used in the analysis and on the probe chosen to inspect the model. In other words, localisation is not a fixed property of the model but an observation shaped by method.

Because the abstract is truncated, the specific circuits and experimental details are not available from the source. The main takeaway is a caution for interpretability work: conclusions about where a behaviour lives in a model can change with the tools used to find it.