Ample research shows that social learning is subject to a similarity bias: People are more likely to copy the behavior of those who resemble themselves. Here we hypothesize that in situations where different approaches must be combined to improve a solution, individuals may instead be more open to learning from dissimilar others because they expect valuable solutions from them. We test this expectation in a preregistered experiment where 859 individuals first developed suboptimal approaches to a problem-solving task and were subsequently exposed to the behavior of a demonstrator. The demonstrator pursued an alternative suboptimal approach; and participants could learn to optimize their own approach by combining it with what they observed. A 2x2 experimental design varied whether the demonstrator had shared or dissimilar characteristics, and whether (dis)similarity related to ideological leaning or the outcome of a fictional cognitive style test. Results show that participants optimized more often after exposure to someone with a different versus similar cognitive style. Ideological (dis-)similarity, on the other hand, did not affect optimization. Our study shows that characteristics pointing to differences in perspectives can promote social learning, challenging the common assumption that people learn best from those similar to themselves.