Researchers have introduced LLaDA-UI, a 16.7B-parameter mixture-of-experts, block-wise diffusion vision-language GUI agent. The work addresses an open question in the field: whether diffusion large language models (dLLMs), which are prized for their block-parallel and arbitrary-order generation, can be extended into capable multimodal GUI agents without sacrificing their decoding efficiency.
GUI agents are presented as a natural testbed for this paradigm because they must repeatedly perceive screen states and emit structured, spatially grounded actions in real time — a latency-sensitive setting where parallel decoding could matter.
LLaDA-UI is trained in two stages. First, general multimodal pre-training aligns a native-resolution vision encoder with the LLaDA2.0-mini-base diffusion language backbone. Second, GUI-agent supervised fine-tuning is performed on diverse mobile, desktop, web, and grounding data.
According to the paper, LLaDA-UI substantially outperforms Qwen2.5-VL-7B across widely adopted grounding and navigation benchmarks spanning multiple platforms, and surpasses Qwen3-VL-8B on four of six reported GUI benchmarks.
The authors conclude that these results establish block-wise diffusion as a practical generative paradigm for multimodal GUI agents. The model is published on Hugging Face.