UNDERSTAND · ACT · EVOLVE

Qiming Zhu朱淇铭

I am a master’s student in the School of Artificial Intelligence at The Chinese University of Hong Kong, Shenzhen, advised by Prof. Benyou Wang.

My research focuses on personalized AI agents. I aim to build capable companions for work and everyday life: agents that understand you, act proactively, and continually improve through long-term collaboration.

Personalization. I study long-term memory as a foundation for agents to use interaction history and understand individual preferences, habits, and needs. TierMem is one step in this direction. I also hope to explore how agents can adapt through experience and feedback.

Proactive agency. I study how agents can perceive changing environments, decide when to act, and use OS and application capabilities to act autonomously and reliably on complex real-world tasks. OS-native Agents explores the system support this requires, with ongoing task management as a testbed.

I also contribute to multi-turn interaction evaluation through MTalk-Bench, studying how to reliably assess models’ interaction capabilities.

Let’s explore and build together, taking steps toward the JARVIS we envision.

Qiming Zhu watching the sunset by the sea
Qiming Zhu · 朱淇铭

Selected research

OS-native Agent framework with persistent task sessions, native state monitors, cross-application capabilities, and authorization.
Research project · 2026

OS-native Agents: Managing User Tasks in the Dynamic Mobile World

Qiming Zhu, Shunian Chen, Wentao Ge, Zhengyang Tang, Xiaoguang Li, Pinjia He, Benyou Wang

Exploring OS support for autonomous, proactive, and reliable action in dynamic environments. The framework brings state awareness, cross-application capabilities, and authorization into the agent architecture. DynamicWorld uses ongoing task management to evaluate agents as environments and user needs change.

TierMem architecture linking summary memory to raw evidence with a router and consolidation loop.
COLM 2026

TierMem: Balancing Compressed Memory and Raw Evidence for Long-Horizon Agent Memory

Qiming Zhu, Shunian Chen, Rui Yu, Zhehao Wu, Benyou Wang

Connecting high-level guidance with detailed instructions through adaptive memory granularity. Compressed memory provides an overview and high-level guidance, while raw records retain detailed information and step-by-step instructions. Agents select the memory granularity needed for the current task and load relevant content on demand.

MTalk-Bench overview connecting speech-aware data construction with Arena-style and rubric-based evaluation.
Preprint · 2025

MTalk-Bench: Evaluating Speech-to-Speech Models in Multi-Turn Dialogues via Arena-style and Rubrics Protocols

Yuhao Du, Qianwei Huang, Guo Zhu, Zhanchen Dai, Shunian Chen, Qiming Zhu, Le Pan, Minghao Chen, Yuhao Zhang, Li Zhou, Benyou Wang, Haizhou Li

Evaluating content, vocal cues, and ambient sound in multi-turn spoken interaction. Pairwise comparisons and detailed rubrics help analyze model performance and the reliability of different evaluation methods.

Education

  1. The Chinese University of Hong Kong, Shenzhen

    Research master’s in Artificial Intelligence · School of AI

    Advisor: Prof. Benyou Wang

  2. Xi’an Jiaotong University

    B.Sc. in Mathematics and Applied Mathematics