Elamur Research⌕
☰
home / papers / 2026 / Me-Agent: A Personalized Mobile Agent with Two-Level User Habit Learning for Enhanced Interaction

Me-Agent: A Personalized Mobile Agent with Two-Level User Habit Learning for Enhanced Interaction

Paper · 2026 · Shuoxin Wang, Chang Liu, Gowen Loo et al.

Summary

LLM-driven phone agents usually execute only what the user states explicitly. They struggle when a request leaves out the app ("play some music") or the content ("play my favorite"), and they do not learn from a person's past behavior. Me-Agent addresses this without any fine-tuning by modeling user habits at two levels. At the prompt level, User Preference Learning (UPL) uses a Personal Reward Model to rank candidate rollouts. It then turns comparisons of successful and failed attempts into natural-language experience. At the memory level, Hierarchical Preference Memory (HPM) keeps long-term user preferences apart from app-specific execution knowledge and retrieves each only when needed. To measure personalization, the authors extend FingerTip into User FingerTip, a benchmark of app-ambiguous and content-ambiguous daily-life instructions. Me-Agent beats Mobile-Agent-v2 and Mobile-Agent-E on both personalization and task execution, and it also generalizes to the multi-app E-dataset. Because it changes no model parameters, the approach fits API-based deployment on resource-limited mobile devices.

Key points

  • User FingerTip covers 60 users, 33 apps in 12 functional categories, 300 Type I (app omitted) and 267 Type II (app and content implicit) instructions, with 15 training and 5 test instructions per user.
  • On User FingerTip, Me-Agent reaches an ASA of 1.000 for both instruction types. Its HSS is 0.890, compared with 0.428 for Mobile-Agent-v2.
  • On the E-dataset, the task completion rate is 0.893, up from a 0.639 baseline. Action accuracy is 0.861 and reflection precision is 0.978. Task success rate (0.600) trails Mobile-Agent-E (0.667).
  • Me-Agent needs fewer steps on the E-dataset, 12.53 compared with 18.40 and 20.20 for the baselines.
  • Ablations on a Mobile-Agent-E base show the two modules are complementary. Performance stays stable across GPT-4o, Claude-3.5-Sonnet and Gemini-2.5-Pro backbones.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Tsinghua UniversityBeijing · China333 works · 87,019 citations🎓Beihang UniversityBeijing · China65 works · 8,452 citations🎓Hong Kong Polytechnic UniversityHong Kong · China84 works · 4,956 citations🎓University of Electronic Science and Technology of ChinaChengdu · China31 works · 2,012 citations🎓Yunnan UniversityKunming · China17 works · 40 citations🎓Southeast UniversityNanjing · China28 works · 354 citations🏢KuaishouBeijing · China15 works · 95 citations🎓Chongqing UniversityChongqing · China7 works · 318 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksInformation Systems · 1,437 worksHuman-Computer Interaction · 141 works

Cite

Cite as Shuoxin Wang, Chang Liu, Gowen Loo, Lifan Zheng, Kaiwen Wei, Huanqian Yan, Xinyi Zeng, Jingyuan Zhang and Yu Tian. 2026. Me-Agent: A Personalized Mobile Agent with Two-Level User Habit Learning for Enhanced Interaction. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.1211

@inproceedings{wang2026me,
  title = {{Me-Agent: A Personalized Mobile Agent with Two-Level User Habit Learning for Enhanced Interaction}},
  author = {Shuoxin Wang, Chang Liu, Gowen Loo, Lifan Zheng, Kaiwen Wei, Huanqian Yan, Xinyi Zeng, Jingyuan Zhang and Yu Tian},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1211},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1211}
}

References

1 references, 1 of them with a card of ours. ↗ leads to the work.

  1. The work is supported by the National Natural Science Foundation of China (62506050), China Postdoctoral Science Foundation Funded Project (2024M763867). We are deeply grateful to our advisor Yu Tian for his extraordinary patience with students and his creative insights into research, both of which…

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wang-2026-me-agent-personalized-mobile-agent/