UI-MOPD: Cross-Platform GUI Agent Distillation

• UI-MOPD is a method for continual GUI agent learning that unifies desktop and mobile behaviors using multi-teacher on-policy distillation.
• It leverages the Uni-GUI dataset and platform-specific teachers to prevent behavioral mixing, capability degradation, and catastrophic forgetting.
• Empirical evaluations show that UI-MOPD improves interactive success rates on both desktop and mobile benchmarks while maintaining robust static grounding.

UI-MOPD denotes the method introduced in “UI-MOPD: Multi-Platform On-Policy Distillation for Continual GUI Agent Learning,” a training recipe for a single GUI agent that must operate on both desktop and mobile platforms while avoiding behavioral pattern mixing, platform-specific capability degradation, and catastrophic forgetting. It combines a unified cross-platform dataset named Uni-GUI, two platform-specific teachers, a smaller shared student, and a platform-conditioned, on-policy multi-teacher distillation objective integrated into reinforcement learning (Lian et al., 5 Jul 2026).

1. Definition and problem scope

UI-MOPD is situated in the setting of cross-platform GUI agents: models that receive a natural-language instruction, one or more screenshots of the current interface, and possibly previous actions and reasoning, and must output the next UI action, such as click, type, scroll, or swipe. The central difficulty is that desktop and mobile environments differ not only in screen layouts and resolutions, but also in action primitives and interaction conventions. The same conceptual operation may therefore require distinct platform-native behaviors, such as a desktop window close operation versus a mobile system back action (Lian et al., 5 Jul 2026).

The method is explicitly motivated by two coupled bottlenecks. First, high-quality and executable cross-platform interaction trajectories are scarce, and existing data often suffer from limited platform coverage. Second, different platforms exhibit distinct interaction conventions, making joint or continual training prone to behavioral pattern mixing, platform-specific capability degradation, and catastrophic forgetting. The paper characterizes this as a seesaw effect of capabilities: improving one platform through naive mixed training or continual updates can degrade performance on the other (Lian et al., 5 Jul 2026).

Within that framing, UI-MOPD is defined as a continual-learning-compatible integration mechanism rather than a pure data-collection effort or a static supervised fine-tuning recipe. Its specific claim is to be the first method that incorporates multi-teacher on-policy distillation into continual learning for GUI agents, with dynamic selection of a platform-specific teacher according to the current environment and transfer of platform-specific behavioral priors to a shared policy through platform-conditioned distillation (Lian et al., 5 Jul 2026).

2. Uni-GUI and the cross-platform data substrate

UI-MOPD is built on Uni-GUI, a cross-platform GUI interaction dataset spanning desktop and mobile environments. Desktop data are drawn from OSWorld environments and OpenCUA trajectories, while mobile data are drawn from MobileWorld, AndroidWorld, and OpenMobile trajectories. The appendix-level composition reported for Uni-GUI is approximately: desktop self-collected data with ~95K steps and ~7K trajectories, cleaned OpenCUA with ~13K steps and ~0.8K trajectories, mobile self-collected data with ~17K steps and ~1K trajectories, and cleaned OpenMobile with ~35K steps and ~2.7K trajectories, for a total of ~160K steps and ~11.5K trajectories (Lian et al., 5 Jul 2026).

The collection harness is organized into four stages. Query generation extracts functional points of each environment and synthesizes natural-language queries grounded in those functionalities. Trajectory collection then uses a strong LLM agent to interact with the environment and records step index, screenshot path and resolution, current query or sub-instruction, model reasoning, action, and possibly environment metadata. Trajectory cleaning removes malformed trajectories, trajectories whose actions cannot be mapped to the student’s standardized action space, trajectories longer than 40 steps, and trajectories whose queries are inconsistent with the environment or unsupported by functional points. Quality judgment is then performed with Gemini-3.1-Pro by decomposing the overall task into sub-tasks and keeping only trajectories where all sub-tasks are completed. Finally, post-processing normalizes reasoning into a structured format aligned with Qwen3-VL’s style, normalizes actions into platform-specific tool interfaces, and re-annotates grounding bounding boxes for training and rule-based reward computation (Lian et al., 5 Jul 2026).

The resulting action schemas are platform-specific but standardized. Desktop episodes use a computer_use tool with actions such as key, type, mouse_move, left_click, left_click_drag, right_click, middle_click, double_click, triple_click, scroll, wait, and terminate. Mobile episodes use a mobile_use tool with actions such as click, long_press, swipe, type, answer, system_button, wait, ask_user, and terminate. The paper also defines AndroidControl*, a static evaluation subset with 4,260 step-level records and 781 trajectories normalized to the same mobile action space (Lian et al., 5 Jul 2026).

3. Teacher–student architecture and platform-conditioned routing

The training pipeline has two stages. In Stage 1, platform-specific supervised teachers are obtained by separately fine-tuning Qwen3-VL-32B-Thinking on the desktop and mobile subsets of Uni-GUI. This yields a desktop teacher $\pi_{\mathrm{ref}}^d$ and a mobile teacher $\pi_{\mathrm{ref}}^m$, each preserving native interaction patterns for its platform. In Stage 2, a single shared student, Qwen3-VL-8B-Thinking, is trained with reinforcement learning plus multi-teacher on-policy distillation (Lian et al., 5 Jul 2026).

The critical architectural choice is platform-conditioned teacher routing. For each sampled rollout, the training system already knows whether the environment is desktop or mobile, and the corresponding teacher is selected accordingly. Formally, for sample $i$ with platform label $s_i$,

$$\pi_{\mathrm{ref}}^{(i)}= \begin{cases} \pi_{\mathrm{ref}}^m, & s_i \in \mathcal{S}_{\mathrm{mobile}} \ \pi_{\mathrm{ref}}^d, & s_i \in \mathcal{S}_{\mathrm{desktop}}. \end{cases}$$

The student remains a single shared policy $\pi_\theta$; only the reference teacher changes from sample to sample (Lian et al., 5 Jul 2026).

This routing rule is central to the method’s attempt to prevent averaged, incoherent policies. Desktop trajectories are constrained only by desktop behavioral priors, and mobile trajectories only by mobile behavioral priors. The paper contrasts this with naive mixed supervised fine-tuning and model merging, which push the model toward an average distribution over actions and can thereby lose sharp, platform-specific behavior (Lian et al., 5 Jul 2026).

The paper does not describe explicit platform embeddings or platform-ID conditioning inside the network. Platform-conditioning is instead realized through the tool schema, the prompts and instructions, and, most importantly, the loss function. This suggests that UI-MOPD’s distinctiveness lies less in architectural branching than in the way optimization is conditioned on platform identity (Lian et al., 5 Jul 2026).

4. Reinforcement-learning objective and on-policy distillation

The reward is defined at the action level through structured comparison with trajectory annotations. For each tool call $a$, the paper defines $f_a \in [0,1]$ as the fraction of matched dimensions, including action type, coordinates within target bounding box, correct scroll direction, correct text, and correct key set, depending on the action. The reward is then

$$R(x, y) = \begin{cases} 1.0, & f_a = 1 \ -0.5, & 0 \le f_a < 1 \ -1.0, & \text{unparsable or invalid action.} \end{cases}$$

This yields positive reward for fully correct actions, moderate penalty for partially correct actions, and strong penalty for invalid outputs (Lian et al., 5 Jul 2026).

The policy-gradient term uses a GRPO-style clipped objective with group-relative advantages. For rollout $i$, the advantage is defined relative to the average reward of its prompt group:

$\pi_{\mathrm{ref}}^m$0

The ratio term is

$\pi_{\mathrm{ref}}^m$1

and the clipped policy-gradient contribution is

$\pi_{\mathrm{ref}}^m$2

The reported hyperparameters include asymmetric clipping with values 0.2 / 0.28 / $\pi_{\mathrm{ref}}^m$3 (Lian et al., 5 Jul 2026).

The distillation term is on-policy: teachers provide probabilities only on the states that the student actually visits. For sampled token $\pi_{\mathrm{ref}}^m$4, the paper defines

$\pi_{\mathrm{ref}}^m$5

and uses the K3 estimator

$\pi_{\mathrm{ref}}^m$6

This is described as a nonnegative, unbiased estimate of $\pi_{\mathrm{ref}}^m$7 under samples from $\pi_{\mathrm{ref}}^m$8, with lower variance than direct log-ratio methods; $\pi_{\mathrm{ref}}^m$9 is clamped in practice for numerical stability (Lian et al., 5 Jul 2026).

The mini-batch distillation loss is

$i$0

where $i$1 masks response tokens and $i$2 is an adaptive group-level KL mask. The mask is turned off when the average group reward exceeds a threshold $i$3:

$i$4

The full loss is

$i$5

with KL coefficient $i$6 (Lian et al., 5 Jul 2026).

The implementation details are correspondingly large-scale. Teachers are trained for 1 epoch each, student RL+MOPD for 1 epoch, with 8 rollouts per prompt, a training batch size of 128, a generation batch size of 384, maximum prompt length 8192 tokens, maximum response length 512 tokens, learning rate $i$7, bfloat16 precision, and 64 $i$8 NVIDIA H100 GPUs arranged as 8 nodes $i$9 8 GPUs (Lian et al., 5 Jul 2026).

5. Empirical behavior on interactive and grounding benchmarks

The principal interactive evaluations are OSWorld for desktop tasks and MobileWorld for mobile tasks. UI-MOPD achieves task success rates of 38.2% on OSWorld and 12.0% on MobileWorld. These are higher than the corresponding results for the base Qwen3-VL-8B-Thinking model, which obtains 33.9% and 7.7%, respectively, and higher than the reported mixed-supervision and model-merging baselines (Lian et al., 5 Jul 2026).

Method	OSWorld	MobileWorld
Qwen3-VL-8B-Thinking	33.9%	7.7%
Mixed-SFT	35.0%	6.4%
Model Merge (weight averaging)	36.5%	6.8%
Model Merge (TIES)	36.8%	0%
UI-MOPD	38.2%	12.0%

The teacher–student analysis clarifies why these gains are presented as continual-learning gains rather than merely scale effects. An 8B model fine-tuned only on OSWorld reaches 35.8% on OSWorld but 0% on MobileWorld, while an 8B model fine-tuned only on MobileWorld reaches 35.8% on OSWorld and 12.8% on MobileWorld. By contrast, the 8B student trained with UI-MOPD reaches 38.2% and 12.0%, improving both sides simultaneously. The platform-specific 32B teachers remain stronger in-domain, with the desktop teacher at 46.3% on OSWorld and the mobile teacher at 16.2% on MobileWorld, but the student closes part of that gap while preserving a shared deployment footprint (Lian et al., 5 Jul 2026).

The paper also reports static and grounding evaluations. On AndroidControl*, UI-MOPD reaches 80.05% overall accuracy, compared with 78.73% for the base 8B model and 74.01% for Model Merge. On OSWorld-G, it reaches 52.84%, slightly above the base model’s 52.13%. On ScreenSpotV2, it reaches 90.88%, close to the base model’s 91.27%. On ScreenSpot-Pro, it reaches 43.14%, again close to the base model’s 43.71%. The fine-grained AndroidControl* results include 87.02% action type accuracy, 88.33% target grounding, and 89.98% ancestor grounding (Lian et al., 5 Jul 2026).

These results are used in the paper to support two claims. First, UI-MOPD improves interactive success on both desktop and mobile benchmarks relative to mixed supervised fine-tuning and merge-based integration strategies. Second, it preserves or slightly improves static grounding and GUI understanding, whereas model merging tends to degrade them. A plausible implication is that platform-conditioned on-policy distillation regularizes policy integration without forcing a collapse of lower-level grounding abilities (Lian et al., 5 Jul 2026).

6. Relation to adjacent research and terminological disambiguation

The name UI-MOPD sits close to several other acronyms in recent literature, but it denotes a distinct method. The nearest conceptual relative is “MOPD: Multi-Teacher On-Policy Distillation for Capability Integration in LLM Post-Training,” which proposes multi-teacher on-policy distillation for combining multiple domain RL teachers into a single LLM and explicitly contrasts that paradigm with Mix-RL, Cascade RL, Off-Policy Finetune, and Param-Merge (Ma et al., 29 Jun 2026). UI-MOPD transfers that general multi-teacher, on-policy, policy-space integration idea into the setting of continual GUI agent learning, with routing determined by platform rather than by domains such as math, instruction following, or SWE (Lian et al., 5 Jul 2026).

It should also not be conflated with “MoPD: Mixture-of-Prompts Distillation for Vision-LLMs,” where MoPD denotes a soft-prompt learning method for CLIP-like models that distills from a mixture of hand-crafted prompt teachers into a learnable soft prompt (Chen et al., 2024). Nor does it coincide with the usage of “UI-MOPD” in “Preference-Guided Multi-Objective UI Adaptation,” where the term refers to a UI multi-objective, preference-driven design paradigm for 3D Mixed Reality layout optimization rather than a GUI-agent training objective (Song et al., 23 Sep 2025).

A second relevant neighboring line is “Ferret-UI 2: Mastering Universal User Interface Understanding Across Platforms,” which addresses universal UI understanding across iPhone, Android, iPad, Webpage, and AppleTV through a single shared multimodal model, adaptive N-gridding, and visually grounded task generation (Li et al., 2024). Ferret-UI 2 covers perception, grounding, advanced QA, and next-action prediction, whereas UI-MOPD addresses cross-platform continual policy learning with platform-conditioned reinforcement learning and teacher routing. This suggests a division of labor between universal UI understanding and multi-platform policy integration: the former concerns perception and grounding, while the latter concerns how to preserve and combine executable behavior across heterogeneous environments (Li et al., 2024, Lian et al., 5 Jul 2026).

7. Limitations and future directions

The current instantiation of UI-MOPD is limited to desktop and mobile platforms. The teacher set contains only two teachers, both at 32B scale, and the implementation is tailored to Qwen3-VL Thinking variants. The reward is rule-based and depends on well-annotated trajectories and grounding boxes. The experiments demonstrate two-platform integration, but not multi-phase continual learning over many rounds or many platform additions (Lian et al., 5 Jul 2026).

The paper therefore frames several natural extensions. One direction is scaling to more platforms and UI paradigms, including web, macOS, multi-monitor setups, or AR interfaces. Another is combining UI-MOPD with richer process rewards or planning modules. Further directions include improving data efficiency, developing stronger exploration strategies, and integrating the method with other continual learning techniques such as self-distillation, parameter regularization, modular architectures, replay buffers, or explicit memory mechanisms (Lian et al., 5 Jul 2026).

Taken in that form, UI-MOPD is best understood as a platform-routed policy-integration method for multimodal GUI agents. Its distinguishing claim is not merely that it trains on both desktop and mobile data, but that it does so through on-policy distillation from platform-specific teachers, thereby attempting to preserve platform-native behavioral priors while training a single shared student policy for continual cross-platform use (Lian et al., 5 Jul 2026).