I am a PhD student in the Department of Computer Science at the University of Rochester, advised by Prof. Chenliang Xu. I received my B.Eng. from CS Department, Huazhong University of Science and Technology in 2022. In my undergrad, I worked with Prof. Kun He and Mr. Xiaosen Wang at HUST on adversarial machine learning. I also work closely with Prof. Yijie Peng at PKU on gradient estimation (Zeroth-Order optimization), Prof. Xiao-Yang Liu at RPI/Columbia on high-performance quantum and tensor computation, and Dr. Xiaodong Liu at Microsoft Research on efficient LLM. I am good at playing Erhu and familiar with Violin. Welcome to reach out to chat:)

Currently, I mainly work on efficient and reliable AI, ranging from classical deep learning models to LLM, for my Ph.D. degree.

Email  /  Google Scholar  /  Github

We propose a parameter-efficient method to post-train the LLM to improve the long-context reasoning ability under limited memory.

DRIFT transfers reasoning from DeepSeek-R1 into QwenVL via gradient-space guidance, improving multimodal reasoning without destabilizing alignment or expensive RL.

We propose a bag of tricks to boost the adversarial transferability of ViT-based attacks.

We propose a novel method to boost audio-visual nerf.

We propose a novel method to unlearn the CLIP on a subgroup of images.

We propose a new benchmark specifically designed to evaluate the video composition understanding capabilities of MLLMs.

We propose a powerful audio-visual adversarial attack and adversarial training defense method.

We propose to allocate the optimal number of queries during the forward-only training to balance estimation accuracy and computational efficiency.

We prune the visual-related computation in multiple MLLMs to accelerate the inference.

We provide early theoretical insights that serve as a roadmap for advancing model ensemble adversarial attack.

We empirically investigate the quantity bias in CLIP.

To alleviate the problem of hallucinations, we propose the Differentiated Beam Decoding (DBD), along with a reliable new set of evaluation metrics: CLIP-Precision, CLIP-Recall, and CLIP-F1.

We propose a method of grouping and pruning similar experts to improve the model's parameter efficienc

This survey provides a detailed overview of the recent advancements in video understanding harnessing the power of LLMs.

We propose a novel method, namely DII (decomposition, identification, and interpretation), to debug the multi-bias of the models.

We propose a novel method, namely L2T (learn to transform), to boost the adversarial transferability.

We explore the potential of Likelihood ratio method for gradient estimation and train multi-architectures of NN without back-propagation.

We propose a bag of novel tricks to boost the adversarial transferability among different models.

We propose to treat the word substitution as a continuous perturbation on the word embedding representation for better robustness.

We propose a novel input transformation based attack, called Structure Invariant Transformation (SIA), which applies a random image transformation onto each image block to craft a set of diverse images for gradient calculation.

We present high-performance tensor-train primitives using GPU tensor cores and demonstrate three applications.

We propose diversifying the high-level features (DHF) for more transferable adversarial examples.

We quantitatively evaluate the robustness of fairness optimization strategies.

We develop a dozen of massively parallel environments to simulate quantum circuits. We open-source our parallel gym environments and benchmarks.

We propose novel hardware-oriented optimization strategies for tensor learning primitives on GPU tensor cores.

We propose a novel Triangle Attack (TA) to optimize the perturbation by utilizing the geometric information that the longer side is always opposite the larger angle in any triangle.

We propose a new technique to compute the pathwise stochastic gradient estimate with respect to the standard deviation of the Gaussian noise added to each neuron of the ANN.

The template is based on Jon Barron's website.