Real-world categories often exhibit graded structure, yet learners struggle to acquire family resemblance categories compared to unidimensional ones in laboratory studies. We propose that part of this difficulty arises from the binary nature of Traditional Artificial Classification Learning. We introduce Graded Classification Learning, a paradigm integrating category and quality judgments into response and feedback phases of a learning trial. This allows higher fidelity feature space exploration, aligning more with naturalistic learning processes. The ‘graded’ learners showed superior performance, with higher final accuracy and steeper learning curves than the ‘traditional’ learners. While aggregate response patterns appeared similar across conditions, profile analysis revealed an apparent gradedness in the traditional condition that was masked by an overwhelming preference for (bisecting) unidimensional strategies, whereas graded participants mostly exhibited genuine graded responses. These findings suggest traditional binary tasks may inadvertently hinder learning of graded structure and that incorporating quality judgments fosters robust category representations.