![[2602.21858] ProactiveMobile: A Comprehensive Benchmark for Boosting Proactive Intelligence on Mobile Devices](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.21858] ProactiveMobile: A Comprehensive Benchmark for Boosting Proactive Intelligence on Mobile Devices
Summary
The paper introduces ProactiveMobile, a benchmark aimed at enhancing proactive intelligence in mobile devices, addressing the limitations of current reactive models.
Why It Matters
As mobile agents evolve, the shift from reactive to proactive intelligence is crucial for improving user experience. ProactiveMobile provides a structured approach to evaluate and advance this capability, addressing a significant gap in the current research landscape.
Key Takeaways
- Proactive intelligence can significantly enhance mobile agent capabilities.
- ProactiveMobile benchmark includes over 3,660 instances across 14 scenarios for comprehensive evaluation.
- Current models show a lack of proactivity, with the best-performing model achieving only a 19.15% success rate.
- The benchmark facilitates systematic research and development in proactive mobile intelligence.
- Expert audits ensure the quality and reliability of the benchmark data.
Computer Science > Artificial Intelligence arXiv:2602.21858 (cs) [Submitted on 25 Feb 2026] Title:ProactiveMobile: A Comprehensive Benchmark for Boosting Proactive Intelligence on Mobile Devices Authors:Dezhi Kong, Zhengzhao Feng, Qiliang Liang, Hao Wang, Haofei Sun, Changpeng Yang, Yang Li, Peng Zhou, Shuai Nie, Hongzhen Wang, Linfeng Zhou, Hao Jia, Jiaming Xu, Runyu Shi, Ying Huang View a PDF of the paper titled ProactiveMobile: A Comprehensive Benchmark for Boosting Proactive Intelligence on Mobile Devices, by Dezhi Kong and 14 other authors View PDF HTML (experimental) Abstract:Multimodal large language models (MLLMs) have made significant progress in mobile agent development, yet their capabilities are predominantly confined to a reactive paradigm, where they merely execute explicit user commands. The emerging paradigm of proactive intelligence, where agents autonomously anticipate needs and initiate actions, represents the next frontier for mobile agents. However, its development is critically bottlenecked by the lack of benchmarks that can address real-world complexity and enable objective, executable evaluation. To overcome these challenges, we introduce ProactiveMobile, a comprehensive benchmark designed to systematically advance research in this domain. ProactiveMobile formalizes the proactive task as inferring latent user intent across four dimensions of on-device contextual signals and generating an executable function sequence from a comprehensive function poo...