How tasks rewire where neural networks store information

Neural networks don't have fixed internal wiring—the same Transformer or Mamba model concentrates information at different layers depending on the task it solves. Across five architectures and three formal language tasks, researchers show that computational structure (how the problem breaks down) matters more than mathematical properties. Crucially, where you can read information with a probe doesn't reveal where the actual computation happens, challenging common mechanistic interpretability assumptions.