Inclusion AI · Venus Team · Westlake University — Technical Report

LLaDA-UI

Bringing block-wise diffusion to vision–language GUI agents: a 16.7B-parameter MoE diffusion model that perceives screens, reasons in tags, and emits executable actions with block-parallel decoding across mobile, desktop, and web.

Zhangxuan Gu2*, Haoxing Chen1*, Qi Qin1*, Yi Xin1*, Kai Gan1, Lin Liu1, Long Cui1, Xiaomei Wang1, Beitong Zhou2, Yunzhu Zhang2, Zhengwen Zeng2, Changlong Gao2, Weizhi Chen2, Rongchao Zhang2, Haoyuan Wu1, Shuheng Shen2, Changhua Meng2, Weiqiang Wang2†, Jianguo Li1†, Zhenzhong Lan1,3†

1AGI Research Center, Inclusion AI  ·  2Venus Team, Ant Group  ·  3Westlake University

* co-first authors  ·  † technical leads

LLaDA2.0-mini-base backbone 16.7B-parameter MoE Mobile · Desktop · Web · Grounding
ScreenSpot-Pro
52.9
grounding · Qwen3-VL-8B: 52.7
AndroidWorld
53.5
mobile · Qwen3-VL-8B: 47.9
MobileWorld
25.6
mobile · Qwen3-VL-8B: 9.4
WebVoyager
56.9
web · Qwen3-VL-8B: 45.2
API speedup
8.95×
max mean speedup vs Qwen3-VL-8B
See it decode

Watch a real prediction denoise

One verbatim AndroidWorld call, replayed frame by frame. The response starts as <|mask|> tokens and resolves block by block into tagged reasoning and one executable action — the observation is uncropped and the output complete; only line wrapping and typography are changed.

Static Android screenshot: a file manager showing the Podcasts folder with holiday_photos.jpg
Static GUI observation
Task Move holiday_photos.jpg from Podcasts to DCIM in the same Android storage area.

Every frame is the recorded model state at one denoising step — reasoning, action tags, and special tokens included. The final frame emits <action>LongPress(box=(x,y))</action>-style executable output.

Two more complete cases

Two further AndroidWorld denoising sequences, exactly as in the report: the first forms a numeric entry action with Type, the second grounds a list item before producing the Click. Each frame carries its own uncropped observation, task, and full model output, including reasoning, action tags, and special tokens.

Overview

Diffusion language models, put to work on screens

GUI agents must repeatedly perceive screen states and emit structured, spatially grounded actions in real time — a natural testbed for diffusion LLMs, whose block-parallel, arbitrary-order generation promises low latency. LLaDA-UI answers the open question of whether that promise survives contact with interactive screens: it substantially outperforms Qwen2.5-VL-7B and surpasses the similarly scaled Qwen3-VL-8B on four of six reported GUI benchmarks, while keeping the parallel-decoding advantage.

LLaDA-UI Qwen3-VL-8B Qwen3.5-9B Qwen2.5-VL-7B

Each axis is normalized to the best displayed raw score (printed under the axis name), mirroring Figure 1 of the report.

Parallel where it counts Block-wise masked diffusion refines many tokens of the current block per step instead of one at a time — and GUI agents, which must repeatedly perceive screen states and emit structured, spatially grounded actions in real time, are a natural testbed for that latency advantage.
Bidirectional context Masked reconstruction conditions on both directions inside a block — bidirectional contextual modeling and global semantic refinement that a strictly causal autoregressive decoder cannot offer.
Structured output that parses Across 1,849 AndroidWorld steps of an intermediate checkpoint: 98.65% of responses parse, 98.59% satisfy the action schema, and all 1,457 coordinate-bearing actions stay in the valid range.
Highlights

What the report establishes

Four results that together make block-wise diffusion a practical generative paradigm for multimodal GUI agents.

The first practical diffusion GUI agent

Diffusion VLMs such as LLaDA-V and SDAR-VL showed an alternative to autoregressive multimodal generation, and diffusion has been explored for static GUI grounding. LLaDA-UI is the first open-source diffusion vision–language agent that repeatedly perceives changing screens, produces executable actions, and completes long-horizon tasks in dynamic environments like AndroidWorld, MobileWorld, OSWorld, and WebVoyager.

Block-parallel decoding, measured

In a controlled paired evaluation with identical serving hardware and no cache hits, LLaDA-UI answers in 4.8–6.4 s where Qwen3-VL-8B needs 17–53 s — 3.58× / 6.83× / 8.95× mean speedups on web, desktop, and mobile inputs, while generating more tokens per call.

A deliberately simple two-stage recipe

145B tokens of general multimodal pre-training align a native-resolution SigLIP ViT with the LLaDA2.0-mini-base diffusion backbone — competitive with VLMs trained on trillions — then GUI-agent SFT on 6M+ samples from the UI-Venus project adds executable behavior across mobile, desktop, web, and grounding.

Tagged reasoning, executable actions

Every navigation target is a tagged <think>…</think><action>…</action> response with spatial outputs normalized to [0,1000]; grounding returns the target point directly. GUI-agent SFT keeps the block diffusion objective and MoE balancing from pre-training, where mask-token reweighting and complementary masking balance learning across targets of substantially different lengths.

Architecture

Three components, one denoising decoder

GUI observations from web, mobile, and desktop environments are encoded at native resolution and interleaved with task and interaction-history tokens; the block-wise diffusion decoder progressively denoises masked output positions into structured reasoning and executable actions.

The LLaDA-UI framework: multi-resolution GUI observations from web, mobile, and desktop pass through a native-resolution vision encoder into the LLaDA 2.0 block-wise diffusion LM decoder, which denoises mask tokens into tagged reasoning and executable actions.
The framework of LLaDA-UI. GUI observations from web, mobile, and desktop environments (1024×768, 1080×2400, and 1920×1080 in the figure) are encoded at native resolution and interleaved with task and interaction-history tokens; the LLaDA2.0 block-wise diffusion decoder progressively denoises masked output tokens into <think>…</think> <action>…</action> responses.
Backbone

Diffusion language model

LLaDA2.0-mini-base — a 16.7B-parameter MoE diffusion LLM following an architecture similar to Ling2.0. Block-wise masked diffusion lets multiple tokens within the current block be refined in parallel instead of one at a time.

MoE · 16.7B total parameters
Perception

Vision encoder

A SigLIP-initialized ViT that ingests screenshots at native resolution and uses 2D rotary positional embeddings to capture the spatial layout that grounding decisions depend on.

native resolution · 2D RoPE
Bridge

Vision–language projector

Groups every four spatially adjacent visual features, then projects them through a two-layer MLP into the LLM embedding space — substantially reducing computational overhead while flexibly compressing image feature sequences of varying lengths.

4× spatial merge · 2-layer MLP
Training

Align first, then act

The vision encoder and the diffusion backbone are pre-trained independently, so stage one aligns their representation spaces on general multimodal data (145B tokens, three-phase curriculum); stage two teaches executable screen behavior with GUI-agent supervised fine-tuning.

S0 · Alignment

Vision–language alignment

High-quality image–caption pairs and visual knowledge collections teach the ViT to produce representations the language model can use — alignment before any full-parameter training.

  • 5B tokens · 16K sequence
  • trains the projector only
S1 · Perception

Perception enhancement

All components unfreeze for joint end-to-end training on OCR, grounding, counting, interleaved image–text, and pure text that preserves the backbone's language ability.

  • 70B tokens · 16K sequence
  • all parameters trainable
S2 · Knowledge

Multi-task pre-training

Reasoning-intensive data lands last: multimodal VQA, multimodal mathematics, and agent-based tasks that build deeper connections between the visual and linguistic modalities.

  • 70B tokens · 16K sequence
  • all parameters trainable

GUI-agent SFT: the data

6M+ samples after conversion, filtering, and deduplication, accumulated through the UI-Venus project across four complementary domains: mobile navigation spanning 100+ Chinese and 70+ English apps, desktop and web data covering cross-application and browser workflows, and grounding data connecting language instructions to visual targets.

GUI-agent SFT: the recipe

Initialized from the multimodal foundation and fine-tuned end to end for three epochs at peak LR 1×10−5 (cosine decay, 3% warmup), 16K max sequence length, mixed precision, micro-batch size one per GPU, ~12.8M max image pixels, on 32–64 H100 GPUs — retaining the block diffusion objective and MoE balancing from pre-training.

The GUI data-generation pipeline: source collection, task seed pool, capability construction, and task generation feed a multi-domain task pool executed on mobile (ADB), web (Playwright), and desktop (VM/PyAutoGUI) environments, producing filtered GUI trajectories.
The GUI data-generation pipeline. Task posing is followed by sub-capability decomposition, capability composition, execution in instrumented mobile, desktop, and web environments, and validation into training trajectories.
Response format — navigation targets are tagged, grounding returns a point
<think> visible reasoning over the task, history, and current screen </think>
<action>Click(box=(x,y))</action>

// available actions (mobile shown; desktop adds RightClick/Hotkey, web adds Scroll/Launch/Hover/Hotkey)
Click · DoubleClick · LongPress · Drag · Swipe · Type(content='')
LaunchApp(app='') · Wait() · CallUser(content='') · GetScreenshot()
PressBack() · PressHome() · PressEnter() · PressRecent()
Answer(content='') · Finished(content='')

// all coordinates normalized to (0,0)–(999,999)
// grounding answers directly: [x,y], or [-1,-1] if infeasible
Block diffusion objective Cross-entropy on masked positions with the diffusion-derived time weight, block by block — the same loss carries from pre-training through SFT.
Mask-token reweighting Inverse square-root weighting over active masked targets, so long responses can't drown out short ones (and vice versa) in the gradient.
Complementary masking Each target yields two corrupted views with inverse masks: every token is observed once and predicted once, cutting token-level sampling bias.
Auxiliary-loss-free MoE balancing RMS-normalized bias updates and routing gates scaled by 2.5 keep expert load uniform without an auxiliary loss term, improving numerical stability during training.
Current-image-only evaluation On MobileWorld, OSWorld, and WebVoyager each request carries exactly one image — the current screenshot — plus the latest raw <think><action> turn; historical screenshots are never replayed (MobileWorld keeps no environment-side history, OSWorld keeps one turn, WebVoyager places the task in the system prompt).
Pre-training stack Multimodal pre-training runs on dFactory, built on VeOmni: distributed multimodal loading, native-resolution processing, boundary-preserving sequence packing, and MoE training.
Results

Benchmark results

Grounding and navigation after GUI-agent SFT, the multimodal foundation before it, paired latency, and the failure analysis — every number from the technical report, with full comparison tables under each chart group.

Static grounding (point-in-box accuracy) and end-to-end navigation success in executable environments. LLaDA-UI exceeds Qwen2.5-VL-7B on every reported benchmark and is stronger than Qwen3-VL-8B on ScreenSpot-Pro, AndroidWorld, MobileWorld, and WebVoyager; ScreenSpot-V2 is close and OSWorld-Verified remains the main gap. Spatial outputs are normalized to [0,1000] and executed by each benchmark runner; both the standalone Hugging Face inference path and the OpenAI-compatible SGLang serving path (two data-parallel workers on a two-GPU host) were validated end to end.

Full comparison table — Grounding & Navigation
Cross-platform traces

Three complete, successful trajectories

Recorded end to end on WebVoyager, OSWorld, and MobileWorld. Every frame pairs the current GUI observation with the model's verbatim <think> and <action> output for that step — step through them or let them play.

Constrained flight search

WebVoyager · 20 steps

Long-horizon constraint tracking on the web: find the one-way Calgary→New York flight on a fixed date with the lowest CO₂ emissions, keeping the constraint in mind across twenty steps of forms, filters, and result lists.

Calc table into Writer

OSWorld · 10 steps

Formatted content transfer between desktop applications: copy a formatted LibreOffice Calc table into Writer and save the resulting document.

Email to alarm

MobileWorld · 12 steps

Cross-application information use on mobile: read an event time from an email, then set an alarm one hour earlier in the clock app.

Takeaway & limits

A viable new direction — stated with its caveats

As the first open-source diffusion GUI agent, LLaDA-UI demonstrates that diffusion language models offer a viable new direction for vision–language agents. The report is equally explicit about where it still falls short:

Long-horizon loops The agent can repeat actions and get stuck in unproductive UI states on long tasks — a pattern also common in GUI agents below the 100B-parameter scale.
High-resolution perception Grounding degrades on very large screenshots with small targets, exposed most clearly by OSWorld's desktop environments.
Decoding sensitivity Small changes to block size and denoising steps can materially alter both task success and latency; decoding robustness is a main direction for future work.
Citation

BibTeX

If you find LLaDA-UI useful, please cite the technical report. arXiv SOON

@article{lladaui2026,
  title   = {LLaDA-UI: Bringing Block-wise Diffusion to Vision-Language GUI Agents},
  author  = {Gu, Zhangxuan and Chen, Haoxing and Qin, Qi and Xin, Yi and Gan, Kai and
             Liu, Lin and Cui, Long and Wang, Xiaomei and Zhou, Beitong and Zhang, Yunzhu and
             Zeng, Zhengwen and Gao, Changlong and Chen, Weizhi and Zhang, Rongchao and Wu, Haoyuan and
             Shen, Shuheng and Meng, Changhua and Wang, Weiqiang and Li, Jianguo and Lan, Zhenzhong},
  journal = {arXiv preprint arXiv:XXXX.XXXXX},
  year    = {2026}
}