Scaling, Benchmarking, and Reasoning of Vision-Language Agents for Mobile GUI Navigation
The paper discusses advancements in Vision-Language Models (VLMs) for mobile GUI navigation. It introduces HyperTrack, a large dataset for evaluating VLM agents, and GUIEvalKit, a toolkit for benchmarking. The findings indicate that reinforcement-based finetuning is more effective than supervised methods, especially in diverse settings.
- ▪Vision-Language Models (VLMs) have shown rapid progress in mobile GUI navigation.
- ▪HyperTrack is a dataset with over 16,000 real-world tasks across more than 650 Chinese mobile applications.
- ▪Reinforcement-based finetuning consistently outperforms supervised finetuning, particularly in out-of-domain settings.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.27134 (cs) [Submitted on 26 May 2026] Title:Scaling, Benchmarking, and Reasoning of Vision-Language Agents for Mobile GUI Navigation Authors:Heng Qu, Yike Liu, Renren Jin, Wenzong Zhang, Pengzhi Gao, Wei Liu, Jian Luan View a PDF of the paper titled Scaling, Benchmarking, and Reasoning of Vision-Language Agents for Mobile GUI Navigation, by Heng Qu and 6 other authors View PDF Abstract:Vision-Language Models (VLMs) have shown rapid progress in mobile GUI navigation. This paper presents a systematic study of data scaling, benchmarking, and reasoning for VLM-based agents in this domain.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
