This study is intended to explore the influence of partners' L2 proficiency on test-takers' performance and on their interaction patterns in paired oral assessment, and it tries to analyze the underlying causes on the basis of Vygotsky's theory of the zone of proximal development (ZPD) and Storch's patterns of interaction. An experiment including three rounds of tests was carried out with 12 students who were selected out of 60 from a top university in China. The findings showed that the high-level test-takers received the highest scores when paired with middle-level partners; the middle-level test-takers scored the highest and the lowest when paired with high-level partners and low-level partners respectively; the low-level test-takers had basically the same scores when paired with those of the other two different levels. The results could be partly explained by the fact that the higher level partners stimulated the lower level test-takers to the upper limit of ZPD through language interaction, but the mechanism of ZPD might not work when partners' proficiency levels were much higher than the ceiling of test-takers' ZPD. Besides, test-takers' performance was mediated by pair interaction patterns. The implications of the study not only deliver suggestions for fairness of paired oral assessment but also provide alternatives for test-takers to achieve higher scores in a paired speaking test by choosing partners of different L2 proficiency levels.