Eye movement recordings from humans indicated that brief exposures (200 ms) to horizontal disparity steps applied to large random-dot patterns elicit horizontal vergence at short latencies (80.9 +/- 3.9 ms, mean +/- SD; n = 7). Disparity tuning curves, describing the dependence of the initial vergence responses (measured over the period 90-157 ms after the step) on the magnitude of the steps, resembled the derivative of a Gaussian, with nonzero asymptotes and a roughly linear servo region that extended only a degree or two on either side of zero disparity. Responses showed transient postsaccadic enhancement: disparity steps applied in the immediate wake of saccadic eye movements yielded higher vergence accelerations than did the same steps applied some time later (mean time constant of the decay, 200 ms). This enhancement seemed to be dependent, at least in part, on the visual reafference associated with the prior saccade because similar enhancement was observed when the disparity steps were applied in the wake of saccadelike shifts of the textured pattern. Vertical vergence responses to vertical disparity steps were qualitatively similar: latencies were longer (on average, by 3 ms), disparity tuning curves had the same general form but were narrower (by approximately 20%), and their peak-to-peak amplitudes were smaller (by approximately 70%). Initial vergence responses usually had directional errors (orthogonal components) with a very systematic dependence on step size that often approximated an exponential decay to a nonzero asymptote (mean space constant +/- SD, 1.18 +/- 0.66 degrees ). Based on the asymptotes of these orthogonal responses, horizontal errors (with vertical steps) were on average more than three times greater than vertical errors (with horizontal steps). Disparity steps >7 degrees generated "default" responses that were independent of the direction of the step, idiosyncratic, and generally had both horizontal and vertical components. We suggest that the responses depend on detectors that sense local disparity matches, and that orthogonal and "default" responses result from globally "false" matches. Recordings from three monkeys, using identical disparity stimuli, confirmed that monkeys also show short-latency disparity vergence responses (latency approximately 25 ms shorter than that of humans), and further indicated that these responses show all of the major features seen in humans, the differences between the two species being solely quantitative. Based on these data and those of others implying that foveal images normally take precedence, we suggest that the mechanisms under study here ordinarily serve to correct small vergence errors, automatically, especially after saccades.