Using must‐have and can‐have features to improve conceptual learning

Williams et al. (2025) taught adults a new concept using picture sets. One group saw only correct examples. Two other groups also saw nonexamples mixed in. The twist: some nonexamples looked almost like the real thing, others looked very different.

The team then tested who could spot the concept in brand-new pictures. They wanted to know if near-miss nonexamples sharpen learning better than far-miss ones.

People who trained with similar nonexamples scored highest on new tests. They picked the concept faster and made fewer mistakes. The examples-only group learned, but their learning did not travel as well to new pictures.

Dissimilar nonexamples helped a little, but not as much as the almost-right ones.

The result lines up with Feinstein et al. (1988). That paper told us to build generalization into training, not tack it on later. Williams shows one clear way: pick nonexamples that sit right on the category border.

It also extends Granerud et al. (2021). Preschoolers formed equivalence classes after naming games. Williams adds a new lever for older learners: fine-tune perceptual similarity instead of adding naming.

Cooper et al. (1990) found equivalence can emerge without extra rewards, but often stays weak. Williams gives a method to make those relations stronger and more general without touching the consequence schedule.

Next time you teach a client to sort ‘safe’ vs ‘dangerous’ items or ‘food’ vs ‘non-food’, slip in look-alikes. A plastic apple that almost matches the real fruit can do more teaching than a rubber ball. One quick change, cleaner concept, fewer errors later.