This paper presents two studies in which a peer-assisted learning condition was compared to an individual learning condition. The first study used the paired-associates learning task and the second study used an incrementally more complex task-the remote associate test. Participants in the peer-assisted learning condition worked in groups of four. They had to solve a given problem individually and give a first answer before being able to request to see their peers' solutions; then, a second answer was issued. After six sessions of peer-assisted practice, a final individual test was administered. Peer interaction was found to benefit learning in both studies but the benefit transferred to the final test only in the second study. Fine-grained behavioral analyses and computational modeling suggested that the benefits of peer interaction were (partially) offset by its costs, particularly increased cognitive load and error exposure. Overall, the superiority of peer-assisted learning over individual learning was more pronounced in the more complex task and for the more difficult problems in that task.