UI-MOPD Enables Cross‑Platform GUI Agents to Retain Skills Without Forgetting

GUI agents face a fundamental dilemma: desktop interaction relies on mouse clicks and drags, while mobile interaction uses touch gestures, creating two incompatible "dialects". Training a single model to master both often results in behavior collapse (35.0% desktop, 6.4% mobile) or catastrophic forgetting (mobile success drops to 0% when using TIES merging).

Problem Analysis

The core issue is that compressing heterogeneous interaction signals into a shared parameter set leads to behavior collapse and catastrophic forgetting, classic challenges in continual learning amplified for GUI agents.

UI‑MOPD Solution

UI‑MOPD (Multi‑Platform On‑Policy Distillation) avoids forcing a model to memorize both rule sets. Instead, it equips the model with dedicated "teachers" for each platform during real‑time learning.

Stage 1 – Teacher Distillation : Using Qwen3‑VL‑32B‑Thinking as a base, separate SFT on desktop and mobile data produces a Desktop Teacher and a Mobile Teacher .

Stage 2 – Online Policy Distillation (MOPD) : An 8B student model explores environments on‑policy. Platform routing directs each trajectory to the corresponding teacher, providing platform‑specific behavior anchors.

teacher = DesktopTeacher if platform == "desktop"
teacher = MobileTeacher if platform == "mobile"

The KL divergence, traditionally a constraint, is turned into an active distillation signal:

KL(π_student ‖ π_teacher) ≈ E[ r - 1 - log r ], r = π_teacher / π_student

Structured rewards (e.g., +1.0 for correct actions, -0.5 for format errors, -1.0 for wrong actions) guide policy optimization:

reward = +1.0 (correct) / -0.5 (format error) / -1.0 (wrong)

Data and Training Infrastructure

The Uni‑GUI data pipeline collects 110k desktop steps and 50k mobile steps, filters them, and yields ~10k high‑quality trajectories (desktop via Kimi‑K2.6, mobile via Gemini‑3.1‑Pro). Training runs on 64 × H100 GPUs.

Results

On the OSWorld (desktop) and MobileWorld (mobile) benchmarks, UI‑MOPD outperforms mixed‑SFT, model merging, and larger models. Compared to the base model, UI‑MOPD improves desktop performance by 12.7% and mobile by 55.8%, and the 8B student surpasses a 235B model (≈30× parameters) on both platforms, demonstrating that gains stem from the method rather than sheer compute.

Tables (omitted here) show that UI‑MOPD retains general GUI understanding (e.g., AndroidControl 80.05%) while model‑merging approaches degrade.

Qualitative Example

A mobile task—replying to a food‑related query and attaching a file—illustrates the agent generating a structured thought chain (observe → plan → reason → anticipate) before executing precise clicks and inputs.

Conclusion

UI‑MOPD proves that assigning platform‑specific teachers and using KL as a guiding distillation signal enables continual, cross‑platform learning without the “forgetting” problem, offering a pragmatic path toward truly universal GUI agents.