Nature Reviews Neuroscience, Published online: 28 August 2026; doi:10.1038/s41583-026-01070-0 Neural network models can express computational hypotheses about brain information processing, but their high parametric capacity makes discriminating model alignment to experimental data challenging. In this Review, Kriegeskorte and colleagues discuss methodology for the optimization of stimuli that make models disagree in their predictions of neural and behavioural data and enhance model comparison.