Yuxi Liu

Peking University

Beijing, China

yuxiliu666@stu.pku.edu.cn

Hi! I am Yuxi Liu (刘禹希). I am currently a 3rd-year Ph.D. student in Data Science at the Center for Machine Learning Research, Peking University (2023-2028), under the supervision of Prof. Kun Yuan. I am also a research intern at Alibaba Tongyi Wanxiang Lab (May 2026-present). Prior to this, I obtained my B.S. degree in Mathematics from Wuhan University (2019-2023).

I am currently interested in Multimodal Large Language Models, Video Generation, and Video World Models. I have served as a reviewer for top-tier conferences including ICML, NeurIPS, ICLR, and CVPR.

I’m always open to research discussions and collaborations. Feel free to contact me at yuxiliu666@stu.pku.edu.cn if you are interested.

💼 Experience

Research Intern, Alibaba Tongyi Wanxiang Lab, May 2026 - Present

📖 Education

Ph.D. in Data Science, Peking University, 2023 - 2028
B.S. in Mathematics, Wuhan University, 2019 - 2023

🏆 Honors and Awards

Excellent Research Award, Peking University, 2025
Outstanding Graduate, Wuhan University, 2023
National First Prize (1st in Hubei Province), Contemporary Undergraduate Mathematical Contest in Modeling (CUMCM), 2021
Meritorious Winner (M Prize), Mathematical Contest in Modeling (MCM/ICM), 2021
First Prize (Hubei Province), Chinese Mathematics Competitions (CMC), 2021
Excellent Scholarship for Math Majors (Class I), Wuhan University, 2021
First-class Scholarship, Wuhan University, 2020
Bronze Medal, Chinese Mathematical Olympiad (CMO), 2018
Gold Medal, China Southeast Mathematical Olympiad (CSMO), 2017

latest posts

Sep 19, 2026	SparkDiffusion: Mitigating the High-Sparsity Trap
Sep 17, 2026	CrossDistill: Balancing Quality and Diversity
Sep 08, 2026	RoLA: Rotary-Positioned Low-Rank Linear Attention
May 27, 2026	RoPeSLR: Sparse-Low-Rank Attention for Diffusion Transformers
Jan 26, 2026	CR-Net: Cross-Layer Low-Rank Structure
Jan 14, 2026	Mixture of Distributions Matters
Nov 01, 2025	MISA: Memory-Efficient LLM Optimization

selected publications

SparkDiffusion: Mitigating the High-Sparsity Trap — A Unified Framework for up to 265x Single-GPU Acceleration of Visual Generation

Yuxi Liu^*, Haoyu Li^*, Zekun Zhang^*, and 12 more authors

arXiv preprint, 2026

Abs arXiv Blog Code Website

Video diffusion transformers are expensive because attention dominates long spatiotemporal token sequences. We identify the high-sparsity trap: at extreme attention sparsity, step-local training losses keep decreasing while terminal generation quality stagnates or degrades. The trap is one of supervision: the dominant terminal errors originate in the high-noise structure-generation stage, and terminal-aligned training corrects terminal errors that substantially extended step-local training cannot. This yields a simple staging principle: first adapt the sparse architecture into a coarse prior, then correct the terminal distribution. We instantiate the principle as SparkDiffusion, a unified acceleration framework for visual generation that combines a short sparse warm-up, few-step trajectory-mixed distillation, and FP8 quantization with fused kernels. SparkDiffusion sustains 97% attention sparsity with strong visual quality on long-sequence 720P generation across Wan2.1/Wan2.2 backbones and T2V/I2V tasks, and 90% sparsity on Wan2.1-T2V-1.3B-480P. With 3-step CFG-free inference, SparkDiffusion achieves a 265x end-to-end speedup over the 50-step CFG dense baseline for Wan2.1-T2V-14B-720P on a single RTX 5090 (220x on H100), and denoises a Wan2.1-T2V-1.3B-480P video in 1.3s.
CrossDistill: Balancing Quality and Diversity via Trajectory-Level Hybrid Few-Step Distillation

Yuxi Liu^*, Haoyu Li^*, Yixiang Cai, and 8 more authors

arXiv preprint, 2026

arXiv Blog Code
RoLA: Rotary-Positioned Low-Rank Linear Attention for Efficient Diffusion Transformers

Zekun Zhang^*, Yixiang Cai^*, Yuxi Liu^*†, and 9 more authors

arXiv preprint, 2026

arXiv Blog Code
RoPeSLR: 3D RoPE-driven Sparse-LowRank Attention for Efficient Diffusion Transformers

Yuxi Liu^*, Zekun Zhang^*, Yixiang Cai^*, and 3 more authors

arXiv preprint, 2026

arXiv Blog
Mixture of Distributions Matters: Dynamic Sparse Attention for Efficient Video Diffusion Transformers

Yuxi Liu^*, Yipeng Hu^*, Zekun Zhang^*, and 2 more authors

In International Conference on Machine Learning (ICML), 2026

arXiv Blog Code Website
CR-Net: Scaling Parameter-Efficient Training with Cross-Layer Low-Rank Structure

Boao Kong^*, Junzhu Liang^*, Yuxi Liu^*, and 2 more authors

In International Conference on Learning Representations (ICLR), 2026

arXiv Blog Code Website
MISA: Memory-Efficient LLMs Optimization with Module-wise Importance Sampling

Yuxi Liu^*, Renjia Deng^*, Yutong He, and 3 more authors

In The Conference on Neural Information Processing Systems (NeurIPS), 2025

arXiv Blog Code Website