Figure 7

Accuracy: how the coordinate is written matters, and slots beat writing

Four ways to write a coordinate

All four arms were trained jointly on computer use (AGUVIS steps + Wave-UI clicks), UI boxes and segmentation, on identical data. They are evaluated on held-out AGUVIS steps and on all 1,272 ScreenSpot clicks.

Slots vs vanilla Gemma 4 31B

On a larger, fully clean test set (1,600 AGUVIS steps from rows not trained on)

Four ways: held-out AGUVIS computer-use steps (200) and ScreenSpot clicks (1,272). Slots vs vanilla: the clean AGUVIS test set (1,600 steps; vanilla on a 300-step sample) and ScreenSpot clicks (1,272; vanilla on a 300-click sample). A step counts as correct when the action matches and, for clicks and swipes, the point is within 14% of the screen of the target; a click counts when it lands inside the element.