Humans comprehend novel objects in light of prior knowledge of familiar categories. This knowledge may be represented differently depending on how it is acquired. Whereas learning through classification biases category knowledge toward diagnostic features, learning through inference produces more complete knowledge (Markman & Ross, 2003). The present study explores how these learning tasks affect people's generalizations about categories, and considers why classification and inference tasks lead to distinctive representations. Experiment 1 finds that classification learners used diagnosticity as a basis for generalization, whereas inference learners used overall typicality. Experiment 2 teases apart the attentional and processing components of the learning tasks with a partial inference condition. In terms of knowledge and generalization, partial inference performance was found to align closely with classification learning. This suggests that the differences between classification and inference learning are largely due to attentional allocation. Discussion focuses on implications for category learning, modeling, and real-world learning.