PersonalAlign: Hierarchical Implicit Intent Alignment for Personalized GUI Agents
A new research paper introduces PersonalAlign, a task for GUI agents that requires aligning with users' implicit intents by leveraging long-term user records. The authors present AndroidIntent, a benchmark with 775 user-specific preferences and 215 routines from 20k records, and propose HIM-Agent, a hierarchical intent memory agent. The work addresses the gap between explicit instruction following and real-world deployment where users have vague or omitted preferences.
Key facts
- PersonalAlign is a new agent task for hierarchical implicit intent alignment.
- AndroidIntent benchmark includes 775 user-specific preferences and 215 routines.
- The benchmark is derived from 20,000 long-term user records.
- HIM-Agent is proposed to maintain continuous hierarchical intent memory.
- The paper is published on arXiv with ID 2601.09636.
- The research focuses on personalized GUI agents with long-term user-centric records.
- The task requires agents to resolve omitted preferences in vague instructions.
- Agents must anticipate latent routines by user state for proactive assistance.
Entities
Institutions
- arXiv