We investigate the effects of 'lecture-based' (LBT)--i.e. individual work and rote learning--vs. 'discussion-based' (DBT)--i.e. participative and focused on student-centred learning--teaching styles on the test scores and socio-economic inequality of middle-school students randomly assigned to classes using data from the China Education Panel Survey--a large-scale nationally representative survey. Estimates from unconditional quantile regressions and decompositions based on the recentred influence function suggest that LBT raises scores in mathematics, but the effect is non-linear, as students in the bottom and top quintiles are more likely to benefit from it. In contrast, LBT lowers scores in Chinese and English. Lecture-based teaching also has greater influence on socio-economically advantaged students, resulting in larger inequality within classrooms, especially between top and median students. These effects arise under various robustness checks, implying that (i) teaching styles affect scores and classroom inequality and (ii) they appear to be subject-specific. These results suggest that teaching styles can be used as a tool to influence students' academic performance as well as the socio-economic heterogeneity that they bring to their classrooms.