The GUI Agent That Learns a New Platform Without Forgetting the Old One
Every GUI agent has the same amnesia problem. Teach it to click around on a phone and it gets worse on desktop; teach it desktop and it forgets mobile. UI-MOPD is a clean attack on that, and it's the second paper this week quietly arguing the same thesis: the training method matters more than the model size.
The setup (arXiv 2607.04425, out July 5) has two pieces. First, Uni-GUI, a cross-platform interaction dataset spanning the different worlds a GUI agent has to live in. Second, the actual trick: multi-teacher on-policy distillation folded into continual learning. Instead of one generalist teacher, they keep a platform-specific teacher for each environment, and use platform-conditioned distillation so the student learns a new platform without smearing its old skills together. It's aimed directly at the catastrophic forgetting and behavior-mixing that wreck cross-platform agents.
The numbers are honest about how hard this still is: 38.2% task success on OSWorld, 12.0% on MobileWorld. Nobody's shipping a flawless computer-use agent off this. But the direction is what counts. These are exactly the kind of environments Bespoke Labs just raised $40M to build, and the mechanism, on-policy distillation, is the same family as the harness-and-skill work that's been eating fine-tuning all spring.
Put the week together and a thesis falls out. The next jump in agent reliability isn't a bigger model, it's better training grounds and smarter ways to teach across them. UI-MOPD is one concrete brick in that wall.
Paper: arxiv.org/abs/2607.04425
← Back to all articles
The setup (arXiv 2607.04425, out July 5) has two pieces. First, Uni-GUI, a cross-platform interaction dataset spanning the different worlds a GUI agent has to live in. Second, the actual trick: multi-teacher on-policy distillation folded into continual learning. Instead of one generalist teacher, they keep a platform-specific teacher for each environment, and use platform-conditioned distillation so the student learns a new platform without smearing its old skills together. It's aimed directly at the catastrophic forgetting and behavior-mixing that wreck cross-platform agents.
The numbers are honest about how hard this still is: 38.2% task success on OSWorld, 12.0% on MobileWorld. Nobody's shipping a flawless computer-use agent off this. But the direction is what counts. These are exactly the kind of environments Bespoke Labs just raised $40M to build, and the mechanism, on-policy distillation, is the same family as the harness-and-skill work that's been eating fine-tuning all spring.
Put the week together and a thesis falls out. The next jump in agent reliability isn't a bigger model, it's better training grounds and smarter ways to teach across them. UI-MOPD is one concrete brick in that wall.
Paper: arxiv.org/abs/2607.04425
Comments