UI-Pro: A Hidden Recipe for Building Vision-Language Models for GUI Grounding
Keywords: Vision-language models; GUI understanding; Visual Grounding
TL;DR: This paper develops a vision-language model (VLM) with enhanced UI grounding abilities through strategic fine-tuning and data scaling, achieving high accuracy with just 25% of the parameters of existing models.
Abstract: Building autonomous UI agents that automate user interactions with interfaces has long been a vision in the field of artificial intelligence. Central to these agents is the capability for UI element grounding, which involves accurately locating UI elements (e.g., buttons and links) based on referring expression, such as user intents and functionality descriptions. Developing these agents with robust grounding capabilities using vision-language models (VLMs) offers a promising path forward. However, a practical framework for creating VLMs with strong element grounding capabilities remains under-explored. To address this gap, we conduct systematic experiments within the design space of VLMs to uncover an effective recipe for building VLMs with strong UI element grounding ability. Firstly, we find that fine-tuning with general visual grounding tasks as a warming-up step mitigates the challenges of fine-tuning with downstream UI element grounding data. Next, we explore different fine-tuning sequences of UI grounding training data from various sources and find that a simple-to-complex fine-tuning curriculum can maximize data utility. Moreover, we find that scaling up the size of either the warming-up data or the UI grounding data in downstream fine-tuning significantly enhances UI element grounding accuracy. Lastly, we explore various image feature compression techniques and find that using a convolution-based compressor to compress UI sub-image features significantly enhances the grounding capabilities on high-resolution UI images. Integrating these insights, we successfully develop UI-Pro, an expert VLM that achieves state-of-the-art UI grounding accuracy with fewer parameters across multiple benchmarks. We hope this work serves as a valuable roadmap for researchers in the UI-VLM domain and inspires future research.
Primary Area: foundation or frontier models, including LLMs
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 1832
Loading