UILoop recasts GUI reasoning as a cyclic Screen-UI elements-Action process and releases a 26K-sample UI Comprehension-Bench, reporting state-of-the-art UI understanding and superior GUI reasoning

Proposes the UILoop paradigm, recasting GUI reasoning as a cyclic Screen-UI elements-Action process rather than direct screen-to-action decision-making. In contrast to existing methods that rely on direct screen-based decision-making, this paradigm places UI elements as an explicit intermediate stage in the reasoning chain. A method-level statement from the abstract, which says the cyclic structure lets MLLMs explicitly learn the localization, semantic functions, and practical usage of key UI elements, yielding precise element discovery and interpretable reasoning.

Introduces a UI Comprehension task centered on UI elements, together with three evaluation metrics. Prior GUI reasoning evaluation did not treat mastery of UI elements as a separate and more challenging task. The abstract explicitly calls the task 'more challenging' and states three evaluation metrics, though it does not name them.

Contributes UI Comprehension-Bench, a benchmark of 26K samples, to comprehensively evaluate existing methods' mastery of UI elements. Provides a dedicated benchmark for UI element understanding instead of reusing existing GUI reasoning evaluation setups. The abstract gives the sample size of 26K and states its purpose is to 'comprehensively evaluate existing methods' mastery of UI elements'.

Experiments report that UILoop achieves state-of-the-art UI understanding performance while yielding superior results in GUI reasoning tasks. Within the authors' stated scope, the paradigm improves both UI understanding and downstream GUI reasoning. The abstract summarizes this with 'Extensive experiments demonstrate' and provides no specific numbers, baselines, or dataset list in the abstract.

The work targets researchers and practitioners building multimodal agents that must understand and operate graphical interfaces, and it applies to settings where GUI reasoning is decomposed into screen perception, UI element understanding, and action execution. UILoop is designed so that MLLMs explicitly learn the localization, semantic functions, and practical usage of key UI elements, so its expected gains concentrate on element-level understanding and interpretable reasoning; the 26K-sample UI Comprehension-Bench provides a common entry point for comparing methods. For readers, this means a reusable task definition, evaluation metrics, and benchmark for probing how well their own models master UI elements, and for designing more interpretable GUI agent pipelines.

The abstract does not name or define the three evaluation metrics, nor does it give the specific numbers, comparison baselines, datasets, or ablation settings behind the state-of-the-art claim, so the magnitude and stability of the 'superior results' remain to be confirmed in the body. How the 26K samples of UI Comprehension-Bench were collected, annotated, and split, and on which GUI environments and task types UILoop was validated, are also not stated in the abstract. In addition, the extra computation and latency cost of cyclic Screen-UI elements-Action reasoning, and the accumulation of error from element-level intermediate representations in long-horizon tasks, are worth watching before adopting the paradigm.