MobiWorld: Mobile Action-Conditioned Models

• MobiWorld is a family of research artifacts that benchmark and predict mobile environments using action-conditioned models across GUI, network, and navigation domains.
• It employs diverse modalities—semantic text, images, code, and diffusion-based architectures—to capture future states and guide autonomous decision-making.
• Researchers leverage MobiWorld to achieve improved task success, energy efficiency, and rollout consistency through reproducible evaluation and counterfactual simulation.

MobiWorld, also written as MobileWorld and appearing in related variants such as MobiWM and MWM, denotes a family of 2025–2026 research artifacts concerned with action-conditioned modeling of mobile environments. In the mobile-agent literature, the name is used for a benchmark of autonomous Android agents operating in agent-user interactive and Model Context Protocol (MCP)-augmented settings (Kong et al., 22 Dec 2025), for semantic and multimodal world models that predict future GUI states (Li et al., 16 Dec 2025, Xu et al., 11 May 2026), and, in adjacent lines of work, for world models of mobile wireless networks, mobile traffic extrapolation, and image-goal navigation (Chai et al., 13 Jul 2025, Qi et al., 9 Apr 2026, Yan et al., 8 Mar 2026). The shared technical motif is prospective prediction under actions; the modeled state varies from screenshots and textual GUI semantics to cellular traffic fields and egocentric RGB observations.

1. Nomenclature and domain scope

In arXiv usage, the label refers to multiple distinct research programs rather than a single unified framework. The term is therefore best understood as a homonym spanning mobile-agent benchmarking, GUI world modeling, mobile-network simulation, and robotics-oriented predictive control.

Artifact	Domain	Stated role
MobileWorld (Kong et al., 22 Dec 2025)	Autonomous mobile agents	Benchmarking in agent-user interactive, and MCP-augmented environments
MobileWorld / MobileWorldBench (Li et al., 16 Dec 2025)	Mobile GUI world modeling	Semantic next-state prediction and planning
MobiWorld (Chai et al., 13 Jul 2025)	Mobile wireless network	Generative world model for planning and optimization
MobiWM (Qi et al., 9 Apr 2026)	Mobile traffic extrapolation	Unlimited-horizon rollout for network planning and optimization
MWM (Yan et al., 8 Mar 2026)	Image-goal navigation	Action-conditioned consistent prediction for planning
“How Mobile World Model Guides GUI Agents?” (Xu et al., 11 May 2026)	Mobile GUI guidance	Comparing delta text, full text, diffusion-based images, and renderable code

A common source of confusion is the proximity between benchmark usage and world-model usage. The benchmark-oriented MobileWorld paper is centered on reproducible evaluation of autonomous agents in Android-like environments (Kong et al., 22 Dec 2025), whereas MobileWorldBench and later GUI world-model papers are centered on predicting future interface states as textual, image, or code representations (Li et al., 16 Dec 2025, Xu et al., 11 May 2026). The wireless-network papers use the same or closely related naming for diffusion- or Transformer-based simulators of radio-network dynamics rather than app interfaces (Chai et al., 13 Jul 2025, Qi et al., 9 Apr 2026).

2. MobileWorld as a benchmark for autonomous mobile agents

The most direct benchmark usage appears in "MobileWorld: Benchmarking Autonomous Mobile Agents in Agent-User Interactive, and MCP-Augmented Environments" (Kong et al., 22 Dec 2025). Its stated motivation is that AndroidWorld had become saturated: recent agentic frameworks exceed 90 % success rates, while AndroidWorld also focuses on short-horizon, single-app tasks, assumes fully specified instructions, and lacks any notion of external-tool invocation through MCP. MobileWorld was introduced to bridge this gap while maintaining reproducible evaluation.

The benchmark comprises 201 tasks across 20 applications. Its task categories are GUI-Only Tasks: 116 (57.7 %), Agent–User Interaction Tasks: 45 (22.4 %), and MCP-Augmented Tasks: 40 (19.9 %). Cross-application structure is central: Single-app tasks: 76 (37.8 %), Two-app tasks: 100 (49.8 %), and $\geq$ Three-app tasks: 25 (12.4 %), so 62.2 % of tasks span multiple applications, compared to 9.5 % in AndroidWorld. MobileWorld average completion steps are 27.8, versus 14.3 for AndroidWorld, making the benchmark nearly twice as long on average (Kong et al., 22 Dec 2025).

Reproducibility is preserved through a snapshot-based container environment in which the entire Android ecosystem (AVD + apps + backends) runs inside Docker-in-Docker. Open-source equivalents of commercial apps are self-hosted, including Mattermost $\simeq$ Slack, Mastodon $\simeq$ Twitter, and Mall4Uni $\simeq$ Taobao. AVD snapshots are restored at the start of each task to guarantee identical initial states. Functional verification combines Textual Answer Verification, Backend Database Inspection, Local Storage Inspection, and Application Callbacks. The paper formalizes evaluation with Success Rate

$\text{SR} = \frac{1}{N} \sum_{i=1}^{N} s_i$

and Average completion steps

$$\text{Ave. Steps} = \frac{1}{N} \sum_{i=1}^{N} t_i.$$

The accompanying agentic framework models the task as a POMDP $$(\mathcal{S}, \mathcal{O}, \mathcal{A}, \mathcal{T}, \mathcal{R})$$, where $$\mathcal{R}: \mathcal{S}\times\mathcal{A}\to\{0,1\}$$. It uses a planner–executor architecture in which the planner consumes the goal, screenshot, and history and emits primitive actions. Besides GUI operations and task-control actions, the action space is extended with ask_user(text) and mcp_call(tool_name, params). Click-type actions are first described in natural language and then grounded to pixel coordinates by a vision-language coordinate predictor.

Under a 50-step limit, the best agentic framework, GPT-5 + UI-Ins-7B, achieves 51.7 % overall success, with 54.0 % on GUI-Only, 62.2 % on Agent–User, and 51.6 % on MCP tasks. The best end-to-end model, Doubao-1.5-UI-TARS, achieves 20.9 % overall, with 26.3 % on GUI-Only and 32.4 % on Agent–User. GPT-5 + UI-Ins-7B also records 27.8 average steps, 1.11 queries per interaction task, $\text{UIQ} = 0.40$, and 2.23 MCP calls per task. The failure analysis identifies hallucinating departure locations instead of asking, context overflow when MCP returns pages of text, no long-term memory leading to repeating rename loops, complex logic or numerical errors in e-commerce tasks, and lack of real-time temporal or spatial grounding causing wrong calendar dates (Kong et al., 22 Dec 2025).

3. Semantic and multimodal world modeling for mobile GUIs

A separate line of work uses MobileWorld not as a benchmark of completed tasks but as a corpus and evaluation suite for predicting the next GUI state. "MobileWorldBench: Towards Semantic World Modeling For Mobile Agents" defines a semantic formulation in which state transitions are represented in natural language rather than pixels (Li et al., 16 Dec 2025). The associated MobileWorld dataset contains 1.4 million state-transition samples $(X_t,a_t,X_{t+1})$ drawn from Android in the Wild and Android Control across 22 apps and four broad task categories. Each sample includes a screenshot $\simeq$0, a natural-language action description $\simeq$1, a screenshot $\simeq$2, three candidate text descriptions $\simeq$3 of the semantic change, and eight candidate yes-no QA pairs about $\simeq$4. After automatic VLM-based filtering and, for the finetuning split, human verification, the released data contains 942 K high-quality change descriptions and 543 K filtered yes-no QA pairs (Li et al., 16 Dec 2025).

The core factorization is

$\simeq$5

where $\simeq$6 is the semantic world model. The benchmark has two tasks: Next-State-Generation, scored by a VLM judge on Accuracy, Completeness, and Relevance with

$\simeq$7

and Next-State-QA, scored by simple accuracy. Finetuning Qwen3-VL-8B-Instruct on MobileWorld raises the generation score from 11.84 to 12.39 and QA accuracy from 67.32% to 71.40%. When integrated into the M3A planning framework on AndroidWorld, success rate rises from 46.9% to 50.8% in zero-shot form and to 54.3% in the finetuned setting (Li et al., 16 Dec 2025).

The later paper "How Mobile World Model Guides GUI Agents?" broadens the design space by comparing four future-state modalities under a unified pipeline: Delta Text, Full Text, Diffusion-Based Image, and Renderable Code (Xu et al., 11 May 2026). Source trajectories from AITW, AMEX, AITZ, AndroidControl, and GUI-Odyssey are converted into $\simeq$8 triples and filtered by graph-level deduplication, render fidelity, action logic, and page-noise removal. Claude-4.6-Opus and Gemini-3-Pro relabel retained samples with delta text, full text, and HTML code. On next-state reconstruction, delta text outperforms full text; for example, MobileWorldModel-8B achieves Delta overall 12.74 versus Full 8.37 on a 1–5 scale. Code2Image achieves SoTA with $\simeq$9, $\simeq$0, and overall 85.4, while diffusion images lag in text fidelity with $\simeq$1 and $\simeq$2. In online AndroidWorld evaluation, delta text raises overall success rate for Qwen3-8B from 40.1 to 44.6, Gemini-3-Flash from 50.6 to 66.6, and GPT-5.4 from 56.0 to 63.8; Code2Image helps capable agents in-distribution but can harm OOD tasks (Xu et al., 11 May 2026).

Taken together, these papers rebut the view that GUI world modeling is necessarily pixel-space prediction. Semantic text latents, delta descriptions, and renderable code are all treated as viable predictive states, with different trade-offs between in-distribution fidelity and online OOD robustness (Li et al., 16 Dec 2025, Xu et al., 11 May 2026).

4. MobiWorld as a generative world model for mobile wireless networks

In wireless-network research, "MobiWorld: World Models for Mobile Wireless Network" defines a generative world model for high-fidelity and flexible environment simulation for mobile network planning and optimization (Chai et al., 13 Jul 2025). The model is designed to learn from heterogeneous, multimodal data from sensors, mobile devices, and base stations and to generate both network element-level observations, such as traffic load and user distribution, and system-level performance indicators, such as throughput and energy consumption.

The architecture has four major stages: data input and tokenization, condition alignment, pretraining with a diffusionTransformer backbone, and fine-tuning. Time-series are encoded by 1D convolutional encoders, images by vision encoders, and graph data by graph-partitioning plus GNN modules. Spatio-temporal contexts $\simeq$3, user-behavior profiles $\simeq$4, and network-configuration parameters $\simeq$5 are mapped into a shared embedding space using condition encoders and contrastive alignment. The diffusion backbone models the joint distribution between mobile network data and conditional factors, with a Mix-of-Experts layer

$\simeq$6

and training uses a self-supervised masking-reconstruction loss alongside the diffusion denoising objective (Chai et al., 13 Jul 2025).

Formally, the forward process is

$\simeq$7

and the reverse model is

$\simeq$8

The simplified training loss is

$\simeq$9

Controllability is achieved by conditioning on spatio-temporal contexts, behavioral profiles, and optimization policies. Sampling from the conditional model yields network-element states and system-level metrics consistent with specified controls (Chai et al., 13 Jul 2025).

The paper’s case study considers collaborative energy saving. At each decision epoch, the agent selects

$\simeq$0

where $\simeq$1 is the number of active carriers in cell $\simeq$2 and $\simeq$3 is the fraction of users offloaded to neighbors. The reward is

$\simeq$4

An agent implemented with PPO or MAPPO optimizes the policy using observations and rewards generated by MobiWorld (Chai et al., 13 Jul 2025).

Quantitatively, for short-term forecasting conditioned on historical data plus $\simeq$5, MobiWorld’s mean absolute error is 8% lower than a specialized LSTM predictor. For long-term generation without history, generated sequences match real statistics within 5% on mean, variance, and autocorrelation. Under five different sleep or offload policies, predicted user-side RSRP has an $\simeq$6 of 0.92 against ground-truth ray-tracing measurements. In energy-saving utility $\simeq$7, MobiWorld-PPO obtains 0.68 under Normal Load, compared with 0.54 for Empirical Thresh. and 0.58 for Heuristic Rule; in High Load scenarios the reported improvements are +21%, +22%, and +27% depending on the cap. PPO converges in fewer than 200 iterations, whereas heuristic methods stagnate after 500 (Chai et al., 13 Jul 2025).

5. MobiWM and unlimited-horizon traffic extrapolation

A related but distinct mobile-network strand is "Beyond Static Forecasting: Unleashing the Power of World Models for Mobile Traffic Extrapolation," which proposes MobiWM (Qi et al., 9 Apr 2026). Here the system state is the per-cell traffic load vector $\simeq$8, and the action is $\simeq$9, consisting of each cell’s transmit power, azimuth, mechanical tilt, and electrical tilt. Rather than using an ELBO-style latent-variable model, MobiWM employs a Transformer encoder–decoder with action encoding and directly minimizes MAE or MSE on predicted traffic.

Its multimodal context fusion combines a facility map image modality, POI and OD-flow grid modalities, shared Fourier positional encoding, and learnable gating modules:

$\text{SR} = \frac{1}{N} \sum_{i=1}^{N} s_i$0

with the sum of gated context added to the decoder hidden state. The model supports unlimited-horizon rollout by recursively feeding predicted traffic back as the next historical window, permitting counterfactual simulation over arbitrary future action trajectories (Qi et al., 9 Apr 2026).

Evaluation is conducted on variable-parameter mobile traffic data covering 31,900 cells across 9 districts. Reported metrics are JSD, MAE, and NRMSE. Across Urban/Suburb and Para/Topo settings, MobiWM achieves the lowest JSD in all four settings and the best MAE in 3/4. For example, in Urban-Para the reported values are 0.3341 / 3.063 / 0.6731 for JSD / MAE / NRMSE, and in Urban-Topo they are 0.3242 / 2.893 / 0.7930. In the downstream PPO case study, the steady-state RMSE is 22,369 $\text{SR} = \frac{1}{N} \sum_{i=1}^{N} s_i$1 4,848 for MobiWM, compared with 23,838 $\text{SR} = \frac{1}{N} \sum_{i=1}^{N} s_i$2 4,952 for Static-MobiWM (no $\text{SR} = \frac{1}{N} \sum_{i=1}^{N} s_i$3), 27,703 $\text{SR} = \frac{1}{N} \sum_{i=1}^{N} s_i$4 4,586 for TimeMoE-WM, and 27,892 $\text{SR} = \frac{1}{N} \sum_{i=1}^{N} s_i$5 8,102 for MobiFM-WM (Qi et al., 9 Apr 2026).

This line differs from the diffusion-based wireless-network MobiWorld in both architecture and target variable. The former emphasizes traffic extrapolation under antenna adjustments and unlimited-horizon rollout, whereas the latter emphasizes controllable generation of broader element-level and system-level network observables (Chai et al., 13 Jul 2025, Qi et al., 9 Apr 2026).

6. MWM and action-conditioned consistent prediction for navigation

A further extension of the naming family appears in "MWM: Mobile World Models for Action-Conditioned Consistent Prediction" (Yan et al., 8 Mar 2026). In the detailed description, the system is presented as a mobile world-modeling framework for image-goal navigation. The model operates in a learned latent space of egocentric RGB observations, with an encoder $\text{SR} = \frac{1}{N} \sum_{i=1}^{N} s_i$6, a Conditional Diffusion Transformer denoiser $\text{SR} = \frac{1}{N} \sum_{i=1}^{N} s_i$7, and a decoder $\text{SR} = \frac{1}{N} \sum_{i=1}^{N} s_i$8. Planning scores candidate trajectories by comparing the decoded terminal frame to the goal image via

$\text{SR} = \frac{1}{N} \sum_{i=1}^{N} s_i$9

The training procedure has two stages. Structure Pretraining uses teacher-forcing diffusion and optimizes

$$\text{Ave. Steps} = \frac{1}{N} \sum_{i=1}^{N} t_i.$$0

building a prior on geometry, lighting, and appearance. ACC Post-training then freezes the CDiT backbone and fine-tunes the lightweight AdaLN layers that inject actions and timesteps, using self-forcing multi-step rollouts and a perceptual consistency loss

$$\text{Ave. Steps} = \frac{1}{N} \sum_{i=1}^{N} t_i.$$1

To preserve rollout consistency under few-step inference, the paper introduces Inference-Consistent State Distillation (ICSD), aligning truncated training states with actual DDIM inference endpoints (Yan et al., 8 Mar 2026).

On the SCAND offline dataset, MWM with DDIM 5 improves rollout consistency relative to NWM. Reported LPIPS values are 0.368 at 1 s and 0.495 at 16 s, versus 0.581 and 0.734 for NWM (DDIM 5). FID at 1 s is 81.0 for MWM (DDIM 5), compared with 167.4 for NWM (DDIM 5), while rollout time is 2.3 s for MWM (DDIM 5), versus 2.6 s for NWM (DDIM 5) and 9.6 s for NWM (DDIM 25). In SCAND navigation with CEM planning, MWM reports ATE 1.14 and RPE 0.302, improving on NWM’s 1.28 and 0.33. On the AIRBOT MMK2 real-world platform, aggregated over four goals, MWM reaches Success Rate 0.30 and Navigation Error 2.16 m, compared with 0.20 and 3.18 m for NWM and 0.08 and 2.88 m for NoMaD (Yan et al., 8 Mar 2026).

Although this work is not a GUI or cellular-network benchmark, it belongs to the same broader research tendency: action-conditioned mobile prediction as a substrate for planning. Its emphasis is not semantic abstraction or benchmark reproducibility, but rollout consistency under self-conditioning and efficient few-step diffusion inference.

7. Comparative themes and interpretive issues

Across these usages, one recurrent misconception is to treat MobiWorld as a single benchmark or a single world-model family. The literature instead uses the name for multiple independent artifacts with distinct state spaces, action spaces, objectives, and metrics. MobileWorld the benchmark reports task success, average steps, user queries, and MCP-call behavior (Kong et al., 22 Dec 2025); MobileWorldBench reports judge-based semantic generation scores and QA accuracy (Li et al., 16 Dec 2025); GUI-guidance work reports reconstruction and downstream-agent performance across delta text, code, and diffusion modalities (Xu et al., 11 May 2026); network papers report MAE, $$\text{Ave. Steps} = \frac{1}{N} \sum_{i=1}^{N} t_i.$$2, JSD, NRMSE, utility, and steady-state RMSE (Chai et al., 13 Jul 2025, Qi et al., 9 Apr 2026); navigation work reports LPIPS, FID, ATE, RPE, success rate, and navigation error (Yan et al., 8 Mar 2026). A plausible implication is that cross-paper comparisons of “MobiWorld performance” are not meaningful without domain qualification.

At a higher level, the research family converges on action-conditioned prospective modeling. In autonomous GUI agents, the question is whether future interface states should be represented as text, code, screenshots, or benchmarkable task outcomes (Kong et al., 22 Dec 2025, Li et al., 16 Dec 2025, Xu et al., 11 May 2026). In mobile-network studies, the question is how to simulate counterfactual traffic, coverage, and energy states under policy interventions (Chai et al., 13 Jul 2025, Qi et al., 9 Apr 2026). In navigation, the question is how to preserve consistency under multi-step imagined rollouts with limited diffusion steps (Yan et al., 8 Mar 2026). This suggests a methodological convergence around controllable transition models, but not a convergence toward a single canonical representation.

The term therefore occupies a broad but coherent position in contemporary arXiv literature: it names mobile-domain systems that move beyond reactive prediction toward explicit modeling of future states under interventions. The principal differences lie in what counts as a state, how actions are encoded, which surrogate of reality is learned, and how correctness is operationalized.