gWorld: Generative Visual Code Mobile World Models

Booking.com

Coupang

Demo videos showing gWorld predicting next GUI states from mobile applications.

Abstract

Mobile Graphical User Interface (GUI) World Models (WMs) offer a promising path for improving mobile GUI agent performance at train- and inference-time. However, current approaches face a critical trade-off: text-based WMs sacrifice visual fidelity, while the inability of visual WMs in precise text rendering led to their reliance on slow, complex pipelines dependent on numerous external models.

We propose a novel paradigm: visual world modeling via renderable code generation, where a single Vision-Language Model (VLM) predicts the next GUI state as executable web code that renders to pixels, rather than generating pixels directly. This combines the strengths of both approaches: VLMs retain their linguistic priors for precise text rendering while their pre-training on structured web code enables high-fidelity visual generation.

We introduce gWorld (8B, 32B), the first open-weight visual mobile GUI WMs built on this paradigm, along with a data generation framework that automatically synthesizes code-based training data. In extensive evaluation across 4 in-distribution and 2 out-of-distribution benchmarks, gWorld sets a new pareto frontier in accuracy versus model size, outperforming 8 frontier open-weight models over 50.25× larger.

Key Results

Average Instruction Accuracy (IAcc.) across all six benchmarks. gWorld 8B and 32B achieve a new pareto frontier in terms of model size (log₁₀ scaled). Notably, extremely large models (e.g., Llama 4 402B) do not reach this pareto frontier.

Key Findings:

gWorld outperforms 8 frontier open-weight models up to 50.25× larger
Code-based approach virtually eliminates structural errors (<1% Render Fail)
+45.7% and +27.1% gain in IAcc. over base models Qwen3 VL 8B, 32B
Scaling training data yields predictable gains following a power law

Data scaling laws for mobile world modeling at 8B. Performance follows a power law (R² ≥ 0.94), indicating predictable and non-saturating gains with more training data.

Method

Schematic diagram of our data generation pipeline. We construct VLM world modeling data via three steps: (1) Repurposing offline policy trajectories into transition triplets; (2) Cross-modal relabeling of the ground-truth next state from pixels to renderable web code; and (3) Synthesizing reasoning traces using look-ahead access to the target state.

Three-Step Data Generation:

Repurposing Policy Trajectory: Convert offline policy trajectories {S_t, A_t} into world modeling data {S_t, A_t, S_t+1}
Synthetic Cross-modal Re-labeling: Convert next-state supervision from pixels to renderable web code using a frontier VLM
Reasoning Data with Look-ahead: Generate reasoning traces R_t with access to ground-truth next state

Qualitative Comparison

Action: click at coordinates (802, 394).

Action: click at coordinates (913, 143).

Action: click at coordinates (756, 685). (AndroidWorld)

Action: click at coordinates (819, 549). (AndroidWorld)

Action: TAP at coordinates (857, 421).

MWMBench: Comprehensive Benchmark

We introduce Mobile World Model Bench (MWMBench), a comprehensive benchmark for evaluating world modeling in mobile GUI environments.

Visual World Modeling

Evaluate in native visual modality preserving rich GUI details

Real-world Action Space

Actions in coordinate space, directly compatible with mobile execution

ID + OOD Evaluation

4 in-distribution + 2 out-of-distribution benchmarks

Datasets: AitW, GUIOdyssey, AndroidControl, AMEX (ID) | AndroidWorld, KApps (OOD)

Main Results

Instruction Accuracy (IAcc., %, ↑)

Model	Size	In-Distribution				OOD		Avg
Model	Size	AitW	GUIO	AC	AMEX	AW	KApps	Avg
Qwen-I-E	20B	15.4	13.0	11.7	10.9	13.8	15.7	13.4
Emu3.5	34B	23.4	25.8	27.7	21.7	29.1	26.8	25.8
Llama 4	109B-A17B	47.6	53.1	50.7	49.0	51.0	48.4	50.0
Llama 4	402B-A17B	47.2	55.8	58.6	58.3	54.3	59.9	55.7
Qwen3 VL	8B	21.5	28.2	31.1	33.7	30.8	30.1	29.2
Qwen3 VL	32B	46.8	52.0	53.2	56.9	53.4	52.5	52.5
Qwen3 VL	235B-A22B	36.1	54.7	51.9	51.2	51.1	64.2	51.5
GLM-4.6V	106B	60.9	68.2	74.2	69.5	74.1	57.4	67.4
gWorld (Ours)	8B	68.8	77.2	78.4	82.6	75.0	67.4	74.9
gWorld (Ours)	32B	71.7	81.5	82.9	86.1	79.9	75.7	79.6

Render Fail (%, ↓)

Model	Size	In-Distribution				OOD		Avg
Model	Size	AitW	GUIO	AC	AMEX	AW	KApps	Avg
Qwen-I-E	20B	—	—	—	—	—	—	—
Emu3.5	34B	—	—	—	—	—	—	—
Llama 4	109B-A17B	4.4	1.2	1.0	0.6	2.9	1.8	2.0
Llama 4	402B-A17B	9.4	7.8	8.6	12.6	14.4	2.2	9.2
Qwen3 VL	8B	33.8	51.4	42.8	31.6	42.3	38.8	40.1
Qwen3 VL	32B	11.6	16.0	13.4	3.8	13.1	8.1	11.0
Qwen3 VL	235B-A22B	40.0	27.2	34.2	30.0	30.0	15.4	29.5
GLM-4.6V	106B	2.4	3.8	1.4	1.2	1.9	4.4	2.5
gWorld (Ours)	8B	0.8	1.2	2.6	0.8	2.3	0.8	1.4
gWorld (Ours)	32B	0.6	0.8	0.8	0.4	0.4	0.6	0.6

Similarity (%, ↑)

Model	Size	In-Distribution				OOD		Avg
Model	Size	AitW	GUIO	AC	AMEX	AW	KApps	Avg
Qwen-I-E	20B	60.1	63.8	63.8	64.4	67.9	71.0	65.2
Emu3.5	34B	68.7	68.8	68.6	71.6	74.2	71.2	70.5
Llama 4	109B-A17B	57.9	62.3	61.4	66.9	61.7	57.3	61.2
Llama 4	402B-A17B	58.9	64.0	63.1	68.1	61.8	58.6	62.4
Qwen3 VL	8B	49.9	48.3	53.4	59.2	49.9	50.5	51.8
Qwen3 VL	32B	59.0	62.7	64.1	70.0	61.2	62.8	63.3
Qwen3 VL	235B-A22B	62.9	69.7	68.8	71.7	65.2	67.3	67.6
GLM-4.6V	106B	64.7	72.5	71.4	73.2	72.2	63.7	69.6
gWorld (Ours)	8B	66.3	73.3	72.8	74.3	69.2	66.1	70.3
gWorld (Ours)	32B	67.3	73.7	74.2	75.4	71.6	66.2	71.4

Main mobile world modeling results across three metrics. gWorld 8B and 32B establish a new pareto frontier in IAcc., achieving near-zero render failures while maintaining competitive visual similarity.

BibTeX

@misc{koh2026generativevisualcodemobile, title={Generative Visual Code Mobile World Models}, author={Woosung Koh and Sungjun Han and Segyu Lee and Se-Young Yun and Jamin Shin}, year={2026}, eprint={2602.01576}, archivePrefix={arXiv}, primaryClass={cs.LG}, url={https://arxiv.org/abs/2602.01576}, }