Data-Driven Learning (DDL), in which learners "confront [themselves] directly with the corpus data" (Johns, 2002, p. 108), has shown to be effective in collocation learning in L2 writing. Nevertheless, there have been only few research studies of this type examining the relationship between English proficiency and corpus consultation. The current study intends to fill the gap by investigating how 140 learners of three different levels of English proficiency from Taiwan utilized Corpus of Contemporary American English (COCA) to correct eight different types of collocation errors adapted from their writing. Data was obtained from three aspects: learners' collocation performance, learners' COCA use, and learners' evaluation toward COCA. A mixed-methods approach that included quantitative statistics and qualitative interviews was used. The results showed that even though learners of higher English proficiency performed better when using corpus, learners across all proficiency levels have improved collocation performance by 30%. Nevertheless, even though lower proficiency learners have received the same amount of assistance from corpus consultation, they did not think corpus as helpful as their higher proficiency fellows did. Teachers should not restrict the use of corpus to higher proficiency learners only because lower proficiency can also benefit from corpus use. [For the complete volume of short papers, see ED572005.]