AI summaryⓘ
The authors created KnowU-Bench, a new test for mobile digital assistants that focuses on how well they learn and use a person's preferences by interacting live with them. Unlike earlier tests that only checked if assistants could follow fixed instructions or guess goals, this benchmark makes agents figure out missing information by talking with users and deciding when to ask for permission before acting. They also built a simulated user that behaves realistically to test these abilities. Their experiments showed that even advanced models struggle to guess preferences and choose when to help, revealing that good basic app control doesn't mean the assistant is good at personalizing help. This points out a big challenge in making trustworthy digital helpers.
Personalized mobile agentsUser preference inferenceProactive assistanceGUI interactionLLM-driven user simulatorConsent negotiationPreference elicitationBenchmark evaluationProactive decision chainClaude Sonnet 4.6
Authors
Tongbo Chen, Zhengxi Lu, Zhan Xu, Guocheng Shao, Shaohan Zhao, Fei Tang, Yong Du, Kaitao Song, Yizhou Liu, Yuchen Yan, Wenqi Zhang, Xu Tan, Weiming Lu, Jun Xiao, Yueting Zhuang, Yongliang Shen
Abstract
Personalized mobile agents that infer user preferences and calibrate proactive assistance hold great promise as everyday digital assistants, yet existing benchmarks fail to capture what this requires. Prior work evaluates preference recovery from static histories or intent prediction from fixed contexts. Neither tests whether an agent can elicit missing preferences through interaction, nor whether it can decide when to intervene, seek consent, or remain silent in a live GUI environment. We introduce KnowU-Bench, an online benchmark for personalized mobile agents built on a reproducible Android emulation environment, covering 42 general GUI tasks, 86 personalized tasks, and 64 proactive tasks. Unlike prior work that treats user preferences as static context, KnowU-Bench hides the user profile from the agent and exposes only behavioral logs, forcing genuine preference inference rather than context lookup. To support multi-turn preference elicitation, it instantiates an LLM-driven user simulator grounded in structured profiles, enabling realistic clarification dialogues and proactive consent handling. Beyond personalization, KnowU-Bench provides comprehensive evaluation of the complete proactive decision chain, including grounded GUI execution, consent negotiation, and post-rejection restraint, evaluated through a hybrid protocol combining rule-based verification with LLM-as-a-Judge scoring. Our experiments reveal a striking degradation: agents that excel at explicit task execution fall below 50% under vague instructions requiring user preference inference or intervention calibration, even for frontier models like Claude Sonnet 4.6. The core bottlenecks are not GUI navigation but preference acquisition and intervention calibration, exposing a fundamental gap between competent interface operation and trustworthy personal assistance.