We Spent a Year Building Verifiable Training Tasks for GUI Agents. Here's What We Learned.
RLVR has scaled in math and code. We tried to do the same for GUI agents: the problems we hit, the shortcuts that didn't work, and what actually did.
Read article →