AIAny
Icon for item

UI-MOPD: Multi-Platform On-Policy Distillation for Continual GUI Agent Learning

Trains cross-platform GUI agents by combining a Uni-GUI cross-platform dataset with platform-conditioned multi-teacher on-policy distillation, enabling a shared policy to adapt to new platforms while retaining platform-specific behaviors; suitable for research on continual GUI agent learning and cross-platform adaptation.

Introduction

Cross-platform GUI agents frequently fail because high-quality, executable multi-platform trajectories are scarce and joint training mixes platform-specific behaviors. UI-MOPD addresses this by aligning the supervision distribution with the agent's deployment states: it supplies platform-conditioned, on-policy distilled guidance from platform-specific teachers while training a single shared policy to operate across desktop and mobile environments. The result is a practical compromise between retaining past capabilities and adapting to new platforms.

Key Findings
  • Introduces Uni-GUI, a curated cross-platform GUI interaction dataset to broaden platform coverage and provide executable trajectories; this reduces reliance on single-platform corpora and supplies examples for continual updates.
  • Proposes platform-conditioned multi-teacher on-policy distillation: at runtime the framework selects a platform-specific teacher and distills its behavioral prior into a shared student policy, which reduces behavioral mixing and catastrophic forgetting across platforms.
  • Demonstrates empirical trade-offs: UI-MOPD achieves reported task success rates of 38.2% on OSWorld and 12.0% on MobileWorld, showing measurable retention-plus-adaptation but leaving headroom for absolute performance improvements.
  • Designed to work in a continual-learning setup: the method incrementally incorporates new-platform data while using teacher guidance to preserve previously learned platform-specific skills.
Who it's for and trade-offs

Great fit if you are researching GUI automation agents, continual learning for interactive systems, or cross-platform adaptation techniques and need a reproducible dataset plus a distillation-based training recipe. Look elsewhere if you require high out-of-the-box success rates on complex mobile tasks (reported mobile performance is lower), have extremely limited compute or cannot provide platform-specific teacher models. The approach reduces catastrophic forgetting but depends on the availability and quality of platform-specific teachers and on-policy rollouts for effective distillation.

Where it fits

UI-MOPD sits between behavior-cloning-only GUI agents (which suffer from off-trajectory supervision gaps) and full multi-agent ensembles: it centralizes inference into a single student model while leveraging multiple teacher policies during training to keep platform priors distinct.

Method sketch

The pipeline conditions the student on platform type and alternates supervised fine-tuning on teacher traces with on-policy distillation from dynamically selected teachers during rollouts, letting the student learn to complete tasks from realistic, policy-induced states while preserving platform-specific behaviors.

Information

  • Websitearxiv.org
  • AuthorsNiu Lian, Alan Chen, Zhehao Yu, Chengzhen Duan, Fazhan Liu, Hui Liu, Pei Fu, Jian Luan, Yaowei Wang, Shu-Tao Xia …
  • Published date2026/07/05

More Items

Coordinates visual and audio forensic agents to locate and verify signs of AI-generated video. Its dual-timescale training improves both the investigation strategy and the specialist inspectors.

AI Others2026

Diagnoses whether robotic manipulation failures stem from history-based reasoning, precise physical execution, or their closed-loop coordination. It spans 30 tasks, three difficulty levels, and two scene settings.

Finds that agent evolution often stalls not for lack of better prompts or skills, but because noisy evaluations fail to identify real improvements. The study tests seven methods across six models.