Zeliang Zhang

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 the CS Department, Huazhong University of Science and Technology in 2022. In my undergraduate studies, 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 LLMs.

Currently, I mainly work on efficient and reliable AI, ranging from classical deep learning models to LLMs. I enjoy playing the Erhu and am also familiar with the violin. Feel free to reach out and chat.

Email / Google Scholar / GitHub

News

[5/2025]	I will work as a research intern at the Deep Learning Group of Microsoft Research, Redmond.
[4/2025]	Welcome to join our workshop at ICCV 2025 in Hawaii on Oct. 20. Workshop site.
[5/2024]	I worked as a research intern at the Deep Learning Group of Microsoft Research, Redmond.
[1/2024]	I worked as an Erhu performer at the Traditional Chinese Ensemble Group of Rochester.
[10/2021]	I worked as a research intern at the Machine Learning for Sustainability Group of Microsoft Research Asia, Beijing.

Research

(* indicates equal contribution with random author order. ‡ indicates the project leader.)

	Training Large Reasoning Models Efficiently via Progressive Thought Encoding Zeliang Zhang, Xiaodong Liu, Hao Cheng, Hao Sun, Chenliang Xu, Jianfeng Gao. ICLR, 2026 We propose a parameter-efficient method to post-train LLMs to improve long-context reasoning ability under limited memory.
	DRIFT: Directional Reasoning Injection for Fine-Tuning MLLMs Chao Huang, Zeliang Zhang, Jiang Liu, Ximeng Sun, Jialian Wu, Xiaodong Yu, Ze Wang, Chenliang Xu, Emad Barsoum, Zicheng Liu. arXiv, 2025 DRIFT transfers reasoning from DeepSeek-R1 into QwenVL via gradient-space guidance, improving multimodal reasoning without destabilizing alignment or expensive RL.
	Harnessing the Computation Redundancy in ViTs to Boost Adversarial Transferability Jiani Liu, Zhiyuan Wang, ‡Zeliang Zhang, Chao Huang, Susan Liang, Yunlong Tang, Chenliang Xu. NeurIPS*, 2025 We propose a bag of tricks to boost the adversarial transferability of ViT-based attacks.
	π-AVAS: Can Physics-Integrated Audio-Visual Modeling Boost Neural Acoustic Synthesis? Susan Liang, Chao Huang, Yunlong Tang, Zeliang Zhang, Chenliang Xu. ICCV, 2025 We propose a novel method to boost audio-visual NeRF.
	Targeted Forgetting of Image Subgroups in CLIP Models Zeliang Zhang, Gaowen Liu, Charles Fleming, Ramana Rao Kompella, Chenliang Xu. CVPR, 2025 We propose a novel method to unlearn CLIP on a subgroup of images.
	VidComposition: Can MLLMs Analyze Compositions in Compiled Videos? Yunlong Tang, Junjia Guo, Hang Hua, Susan Liang, Mingqian Feng, Xinyang Li, Rui Mao, Chao Huang, Jing Bi, Zeliang Zhang, Pooyan Fazli, Chenliang Xu. CVPR, 2025 We propose a new benchmark specifically designed to evaluate the video composition understanding capabilities of MLLMs.
	Rethinking Audio-Visual Adversarial Vulnerability from Temporal and Modality Perspectives ‡Zeliang Zhang, Susan Liang, Daiki Shimada, Chenliang Xu. ICLR, 2025 We propose a powerful audio-visual adversarial attack and adversarial training defense method.
	FLOPS: Forward Learning with OPtimal Sampling Tao Ren, Zishi Zhang, Jinyang Jiang, Guanghao Li, Zeliang Zhang, Mingqian Feng, Yijie Peng. ICLR, 2025 We propose to allocate the optimal number of queries during forward-only training to balance estimation accuracy and computational efficiency.
	Treat Visual Tokens as Text? But Your MLLM Only Needs Fewer Efforts to See Zeliang Zhang, Phu Pham, ‡Wentian Zhao, ‡Kun Wan, Yu-Jhe Li, Daniel Miranda, Ajinkya Kale, Chenliang Xu. Preprint, 2024 We prune the visual-related computation in multiple MLLMs to accelerate inference.
	Understanding Model Ensemble in Transferable Adversarial Attack Wei Yao, Zeliang Zhang, Huayi Tang, Yong Liu. ICML, 2025 We provide early theoretical insights that serve as a roadmap for advancing model ensemble adversarial attack.
	Can CLIP Count Stars? An Empirical Study on Quantity Bias in CLIP Zeliang Zhang, Zhuo Liu, Mingqian Feng, Chenliang Xu. Findings of EMNLP, 2024 We empirically investigate the quantity bias in CLIP.
	Do More Details Always Introduce More Hallucinations in LVLM-based Image Captioning? Mingqian Feng, Yunlong Tang, Zeliang Zhang, Chenliang Xu. Preprint, 2024 To alleviate hallucinations, we propose Differentiated Beam Decoding (DBD), along with CLIP-Precision, CLIP-Recall, and CLIP-F1.
	Diversifying the Expert Knowledge for Task-Agnostic Pruning in Sparse Mixture-of-Experts Zeliang Zhang, Xiaodong Liu, Hao Cheng, Chenliang Xu, Jianfeng Gao. ACL Findings, 2025 We propose grouping and pruning similar experts to improve parameter efficiency in sparse MoE models.
	Video Understanding with Large Language Models: A Survey Yunlong Tang, Jing Bi, Siting Xu, Luchuan Song, Susan Liang, Teng Wang, Daoan Zhang, Jie An, Jingyang Lin, Rongyi Zhu, Ali Vosoughi, Chao Huang, Zeliang Zhang, Feng Zheng, Jianguo Zhang, Ping Luo, Jiebo Luo, Chenliang Xu. Technical Report, 2023 This survey provides a detailed overview of recent advancements in video understanding with LLMs.
	Discover Multiple Biased Subgroups in Image Classifiers Zeliang Zhang, Mingqian Feng, Zhiheng Li, Chenliang Xu. CVPR, 2024 We propose DII (decomposition, identification, and interpretation) to debug multiple biases in models.
	Learning to Transform Dynamically for Better Adversarial Transferability Rongyi Zhu, Zeliang Zhang‡*, Susan Liang, Zhuo Liu, Chenliang Xu. CVPR*, 2024 We propose L2T (learn to transform), a novel method to boost adversarial transferability.
	One Forward is Enough for Training Neural Networks via the Likelihood Ratio Method Jinyang Jiang, Zeliang Zhang, Chenliang Xu, Zhaofei Yu, Yijie Peng. ICLR, 2024 We explore the potential of the likelihood ratio method for gradient estimation and train multiple neural architectures without backpropagation.
	Bag of Tricks to Boost the Adversarial Transferability Zeliang Zhang, Wei Yao, Xiaosen Wang Technical Report, 2024 We propose a bag of novel tricks to boost adversarial transferability among different models.
	Random Smooth-based Certified Defense against Text Adversarial Attack Zeliang Zhang, Wei Yao, Susan Liang, Chenliang Xu EACL Findings, 2024 We treat word substitution as a continuous perturbation on word embeddings for better robustness.
	Structure Invariant Transformation for Better Adversarial Transferability Xiaosen Wang, Zeliang Zhang, Jianping Zhang ICCV, 2023 We propose Structure Invariant Transformation (SIA), a novel input transformation attack for more diverse gradients.
	High-performance Tensor-Train Primitives Using GPU Tensor Cores Xiao-Yang Liu, Hao Hong, Zeliang Zhang, Weiqing Tong, Xiaodong Wang, Anwar Walid IEEE Transactions on Computers, 2024 We present high-performance tensor-train primitives using GPU tensor cores and demonstrate three applications.
	Diversifying the High-level Features for Better Adversarial Transferability Zhiyuan Wang, Zeliang Zhang, Siyuan Liang, Xiaosen Wang BMVC, 2023, oral We propose diversifying high-level features for more transferable adversarial examples.
	How Robust is your Fair Model? Exploring the Robustness of Diverse Fairness Strategies Edward Small, Wei Shao, Zeliang Zhang, Peihan Liu, Jeffrey Chan, Kacper Sokol, Flora Salim Data Mining and Knowledge Discovery, 2024 We quantitatively evaluate the robustness of fairness optimization strategies.
	Classical Simulation of Quantum Circuits: Parallel Environments and Benchmark Xiao-Yang Liu, Zeliang Zhang NeurIPS Datasets and Benchmarks Track, 2023 We develop massively parallel environments to simulate quantum circuits and open-source the benchmark suite.
	High-Performance Tensor Learning Primitives Using GPU Tensor Cores Xiao-Yang Liu, Zeliang Zhang, Zhiyuan Wang, Han Lu, Xiaodong Wang, Anwar Walid. IEEE Transactions on Computers, 2022 We propose hardware-oriented optimization strategies for tensor learning primitives on GPU tensor cores.
	Triangle Attack: A Query-efficient Decision-based Adversarial Attack Xiaosen Wang, Zeliang Zhang, Kangheng Tong, Dihong Gong, Kun He, Zhifeng Li, Wei Liu ECCV, 2022 We propose Triangle Attack (TA), which leverages triangle geometry to optimize perturbations efficiently.
	Noise Optimization for Artificial Neural Networks Li Xiao, Zeliang Zhang, Yijie Peng Short paper: CASE, 2022; Long paper: T-ASE, 2024 We propose a new technique to compute pathwise stochastic gradient estimates with respect to neuron noise standard deviations.

Education

	University of Rochester, NY, USA Ph.D. in Computer Science Sep. 2022 - Present Advisor: Chenliang Xu
	Huazhong University of Science and Technology, Wuhan, China B.Eng. in Computer Science and Technology Sep. 2018 - Jun. 2022

Experience

	Microsoft Research, Redmond, US Research Intern, then Part-time Researcher May 2025 - Dec. 2025 Advisor: Xiaodong Liu and Hao Cheng Work on efficient training and inference of reasoning language models.
	Microsoft Research, Redmond, US Research Intern, then Part-time Researcher May 2024 - Nov. 2024 Advisor: Xiaodong Liu and Hao Cheng Work on efficient training and inference of language models.
	Microsoft Research Asia, Beijing, China Research Intern Oct. 2021 - Jun. 2022 Advisor: Xinran Wei Work on high-performance computation of DFT, an important bottleneck in AI-driven material design.
	Columbia University, NYC, US Research Assistant Feb. 2020 - Dec. 2022 Advisor: Xiao-Yang Liu Work on high-performance tensor computation using GPUs and publish two workshop papers: Trillion-Tensor: Trillion-Scale CP Tensor Decomposition (IJCAI 2020 TNRML Workshop) and Parallel TTr1-Tensor: Randomized Compression-based Scheme for Tensor Train Rank-1 Decomposition (NeurIPS 2020 QTNML Workshop).

Projects

	ElegantRL (~ 3k stars! 🚀) Develop the RL-Hamiltonian algorithm to stabilize RL training and publish it as a poster in the NeurIPS 2021 QTNML Workshop.
	TransferAttack One of the main contributors. TransferAttack is a PyTorch framework to boost adversarial transferability for image classification.
	RL for Quantum Circuits One of the main contributors. Open-source benchmark and environments for classical simulation of quantum circuits.