UI-MOPD hits 38.2% on OSWorld via multi-teacher distillation

Cross-platform generalization is the quiet failure mode of the current wave of GUI agents, and a new arXiv technical report titled UI-MOPD is a bet that the fix is more about how you train than how big the model is. When you take an agent that works on a desktop like OSWorld and try to bring it up on a mobile environment, joint or continual training tends to blur behaviors together and, in the authors' own words, produce "behavioral pattern mixing, platform-specific capability degradation, and catastrophic forgetting." That is the ceiling they are trying to lift.

Their approach, which they describe as the first to bring multi-teacher on-policy distillation into continual learning for GUI agents, keeps a separate platform-specific teacher for each environment and swaps between them at training time depending on which platform the agent is currently interacting with. Priors from that teacher get pushed into a shared policy through what they call platform-conditioned distillation. Alongside the method they release Uni-GUI, a cross-platform GUI interaction dataset built to address the scarcity of executable trajectories that span more than one operating system.

The reported numbers, per the technical report, are a 38.2% task success rate on OSWorld and 12.0% on MobileWorld. The OSWorld figure is the eye-catching one. The MobileWorld figure is a useful reminder that mobile automation is still much harder than desktop, and that a headline OSWorld number on its own would over-sell where cross-platform agents actually are today.

The honest caveat is that this is a 25-page technical report, not peer-reviewed work, and it is a single-lab result without independent replication yet. What the paper's abstract does not give you is which base model was distilled, how large it is, how much compute Uni-GUI required to build, or whether the platform-conditioned trick generalises to a third or fourth surface, web browsers most obviously, which is exactly the question a builder would want answered before betting a roadmap on it.

If it holds up, the interesting readers are teams building enterprise agent stacks, where the ability to add a new platform without regressing the last one is worth more than a raw benchmark win. Continual learning with per-platform teachers is a shape you can imagine copying.