A study investigated the variability of language performance on different types of testing task, global versus discrete-focus. Three tests (cloze, multiple-choice, and fill-in-the-blank) were developed to measure learners' knowledge of five verb forms. The tests, containing corresponding items designed to elicit equivalent structures, were administered to nonnative speakers of English grouped by proficiency level and by language background and also administered to a smaller group of native speakers. The results showed a clear pattern of variability, with students performing best on the multiple-choice task and least well on the cloze task, with greater variability at lower proficiency levels and on the more difficult verb structures. Differences in performance also seemed to be closely related to the production versus recognition features of the elicitation task. Analysis of two other factors, language background and error type, suggest a role for first language in performance variability. (MSE)