Kuai (Clara) Yu 俞快

Hi! I’m Kuai. My name means happiness and speed in Chinese, two things I bring to both my research and my daily life. I’m a second year master student in Computer Science from Columbia Engineering. I’m very fortunate to be supervised by Prof. Hod Lipson in Creative Machines Lab from Columbia University. I also work as a research assistant with Prof. Huan Zhang at University of Illinois Urbana-Champaign. Also, I am supervised by Prof. Hao Wang in Intellisys Lab from Stevens Institute of Technology.

My research interests include Web Agents, Trustworthy ML, and Computer Vison.

I am looking for a PhD position in Fall 2026!!

Publications

How do Visual Attributes Influence Web Agents? A Comprehensive Evaluation of User Interface Design Factors

Kuai Yu, Naicheng Yu, Han Wang, Rui Yang, Huan Zhang

Under Review

Paper Link

We propose VAF, a comprehensive evaluation framework that systematically generates CSS-based webpage variants to analyze how visual design elements such as color, layout, and typography influence web agents’ decision-making, revealing human-like visual biases and vulnerabilities in multimodal web agents.

LyTimeT: Towards Robust and Interpretable State-Variable Discovery

Kuai Yu, Crystal Su, Xiang Liu, Judah Goldfeder, Mingyuan Shao, Hod Lipson

2026 IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP26)
Paper Link

We propose LyTimeT, a vision-transformer-based framework that integrates Lyapunov-theoretic regularization with time-series modeling to discover interpretable and robust state variables from visual dynamical systems, enabling stable trajectory prediction and physically consistent variable extraction across diverse temporal domains.

POLAR: Policy-based Layerwise Reinforcement Method for Stealthy Backdoor Attacks in Federated Learning

Kuai Yu, Xiaoyu Wu, Peishen Yan, Yang Hua, Hao Wang, Tao Song, Linshan Jiang, Haibing Guan

Under Review
Paper Link

We propose POLAR, a reinforcement learning-based method that formulates layer-wise selection for efficient and stealthy backdoor attacks in federated learning. POLAR dynamically adapts to defenses and balances attack success rate with stealthiness.

W3FedPOSE: A Practical WEB3.0-Based Federated Learning Framework with Privacy, Ownership, Security, and Efficiency

Peishen Yan, Kuai Yu, Yaozhi Zhang, Yulin Sun, Shuang Liang, Yang Hua, Tao Song, Linshan Jiang, Ningxin Hu, Mohammad Reza Haghighat, Bingsheng He, Haibing Guan

Under Review

We present W3FedPOSE, a practical federated learning framework that integrates Web3.0 technologies. It leverages blockchain and smart contracts for secure data ownership and access control, while introducing an explainable, incentive-driven algorithm for efficient and trustworthy collaboration.

Services

Reviewers for AAAI2026.