Teachers play a central role in shaping how students benefit from peers, yet little is known about how classroom composition affects their attention-allocation decisions. We conduct a large-scale randomized experiment using realistic classroom vignettes to assess how teachers engage with students under varying scenarios and objectives. The presence of a high achiever reduces the likelihood that teachers engage with a low achiever by about 8%, with substantially larger effects when teachers prioritize task success, consistent with convenience-based decision-making. Using administrative data, we show that low achievers perform worse when quasi-randomly assigned to a classroom with an exceptional student.