AI/ ai · gui agents · multimodal · continual learning

UI-MOPD Teaches GUI Agents New Platforms Without Losing Old Skills

A new training method called UI-MOPD uses on-policy distillation to stop GUI agents from forgetting one platform's conventions when learning another.

A research team has built a training framework that lets a single AI agent navigate desktop and mobile interfaces across platforms without the usual amnesia that comes with learning new ones.

The core problem: GUI agents trained on multiple platforms tend to mix up interaction conventions or lose competence on earlier platforms as they learn new ones — a well-documented failure mode called catastrophic forgetting. The researchers address this with UI-MOPD, which pairs each platform with a dedicated "teacher" model and distills that teacher's behavioral patterns into a shared policy only when the agent is actively working in that environment. They also released Uni-GUI, a cross-platform interaction dataset built to support this kind of training. On the OSWorld desktop benchmark the method hit a 38.2% task success rate; on MobileWorld it reached 12.0%.

Those numbers matter because they represent a meaningful step toward agents that can actually switch contexts — from a Linux desktop to an Android phone — without a human retraining them from scratch each time. Most prior GUI agent work optimizes for a single platform, which limits real-world utility considerably.

The 12.0% success rate on MobileWorld is a reminder that "works across platforms" still means "struggles on most tasks" — but that's where the field is, and UI-MOPD at least makes the failure modes more predictable than its predecessors.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →