[Paper Review] PersonalAlign: Hierarchical Implicit Intent Alignment for Personalized GUI Agent with Long-Term User-Centric Records

Yibo Lyu, Gongwei Chen+3 more|arXiv (Cornell University)|Jan 14, 2026

Personal Information Management and User Behavior0 citations

TL;DR

Introduces PersonalAlign with HIM-Agent and AndroidIntent benchmark to enable GUI agents to align with implicit user intents using long-term records, showing improved execution and proactive performance.

ABSTRACT

While GUI agents have shown strong performance under explicit and completion instructions, real-world deployment requires aligning with users' more complex implicit intents. In this work, we highlight Hierarchical Implicit Intent Alignment for Personalized GUI Agent (PersonalAlign), a new agent task that requires agents to leverage long-term user records as persistent context to resolve omitted preferences in vague instructions and anticipate latent routines by user state for proactive assistance. To facilitate this study, we introduce AndroidIntent, a benchmark designed to evaluate agents' ability in resolving vague instructions and providing proactive suggestions through reasoning over long-term user records. We annotated 775 user-specific preferences and 215 routines from 20k long-term records across different users for evaluation. Furthermore, we introduce Hierarchical Intent Memory Agent (HIM-Agent), which maintains a continuously updating personal memory and hierarchically organizes user preferences and routines for personalization. Finally, we evaluate a range of GUI agents on AndroidIntent, including GPT-5, Qwen3-VL, and UI-TARS, further results show that HIM-Agent significantly improves both execution and proactive performance by 15.7% and 7.3%.

Motivation & Objective

• Motivate the need for GUI agents to infer users' implicit intents beyond explicit instructions.
• Propose a hierarchical view of implicit intent to handle preference and routine alignment.
• Create AndroidIntent to annotate long-term user records for evaluation.
• Develop HIM-Agent to maintain and organize long-term memory for personalization.
• Demonstrate improved performance of HIM-Agent on the AndroidIntent benchmark.

Proposed method

• Define PersonalAlign task with three paradigms: Reactive, Preference, and Routine alignment.
• Construct AndroidIntent, a long-term, user-centric GUI benchmark with hierarchical filtering for annotation.
• Propose HIM-Agent with a Streaming Aggregation Module to incrementally update memory.
• Develop Execution-based Preference Filter and State-based Routine Filter to form hierarchical memory for preferences and routines.
• Combine dense embeddings with sparse Jaccard and use DTW for action trajectory similarity in memory updates.
• Evaluate across multiple GUI agents (GPT-5, Qwen3-VL, UI-TARS, etc.) and show performance gains.

Experimental results

Research questions

• RQ1 How can GUI agents infer and align with users' implicit preferences from long-term records when instructions are vague?
• RQ2 How can hierarchical memory structures and streaming updates support preference and routine intent in GUI agents?
• RQ3 To what extent does personalized implicit-intent alignment improve reactive execution and proactive assistance in GUI tasks?

Key findings

• HIM-Agent significantly improves execution and proactive performance over baselines by 15.7% and 7.3%, respectively.
• AndroidIntent provides annotated ground-truth for 775 preference intents and 215 routine intents from 20k long-term records across 91 users.
• A Streaming Aggregation Module and hierarchical memory (Preference vs Routine) enable stable, scalable personalization.
• Ablation studies show all components of the Execution-based Preference Filter contribute to performance gains, with full module yielding notable CER improvements.
• Proactive evaluation demonstrates better balance between intent alignment and false alarms across open- and closed-source GUI agents.