The structure combination approach provides the best account of our data. We developing a model that accounts for across-object and across-feature generalization, including cases where both kinds of generalization must work in tandem. Our model simultaneously draws on taxonomic relationships between objects and causal relationships between features, and our experiments confirm that people are able to combine these two kinds of information. knowledge captured by this model falls well short of the complexity of commonsense knowledge. Accounts of generalization will eventually need to grapple with this complexity, and future studies of inductive reasoning will need to explore how many different pieces of knowledge are integrated and composed.