Yuxiang Liu

Yuxiang Liu

Tianjin University

Hi, I’m Yuxiang Liu (刘宇翔). Welcome to my homepage!

I am an undergraduate student at the College of Intelligence and Computing, Tianjin University (TJU).

My research interests include 3D Reconstruction, multimodal video-audio generation, and egocentric spatiotemporal reasoning.

My work has been accepted by ICML 2026, and my team won first place in the Skeleton Tracking Challenge at the CVPR 2025 Global 3D Human Poses (G3P) Workshop. I have also received the National Scholarship for two consecutive years.

My Email: lyx1021@tju.edu.cn

🔥 News

  • 2026.09:  🚀 Released EgoTSR++, a framework for egocentric spatiotemporal reasoning and task-progress understanding.
  • 2026.08:  🚀 Released TurboT2VA, our framework for fast joint text-to-video-audio generation, achieving 54.67× generator-only speedup at high resolution on a single NVIDIA H20.
  • 2026.05:  🎉 Paper accepted by ICML 2026: EgoTSR — Evolving Ego-Centric Task-Oriented Spatiotemporal Reasoning via Curriculum Learning.
  • 2025.12:  ⭐ Awarded the National Scholarship for the academic year 2024-2025.
  • 2025.06:  🏆 Won first place in the Skeleton Tracking Challenge at the CVPR 2025 G3P Workshop.
  • 2024.12:  ⭐ Awarded the National Scholarship for the academic year 2023-2024.

📝 Publications

(* co-first author, † corresponding author)

  • EgoTSR++: Egocentric Spatiotemporal Reasoning for Task Progress Understanding
    Xiaoda Yang*, Can Wang*, Yuxiang Liu, Pengfei Zhou, Jianwen Lou, Shuicheng Yan†
    arXiv preprint, 2026  [Paper]
    • Introduced a framework for diagnosing and improving order-robust task-progress understanding in egocentric scenarios.
    • Built SpatialLogic-Bench to evaluate physical state pairs in both original and order-swapped presentations, exposing input-order bias in representative VLMs.
    • Combined bidirectional supervision, failure-aware data, and a progressive CoT-to-Tag curriculum, achieving 92.4% long-horizon accuracy and improving non-monotonic trajectory understanding.
  • TurboT2VA: Fast Large-Scale Text-to-Video-Audio Generation via Score-Regularized Consistency Distillation
    Xiaoda Yang*, Yuxiang Liu*, Kaiwen Zheng, Yuan Liu, Yibo Lai, Shengpeng Ji, Kai Jiang, Jianfei Chen, Shan Yang, Sen Liang, Xiaobin Hu, Shuicheng Yan, Jintao Zhang†, Jun Zhu†, Zhou Zhao†
    arXiv preprint, 2026  [Paper | Code & Demos]
    • Proposed TurboT2VA, a score-regularized consistency distillation framework that accelerates a 19B-parameter joint video-audio model while preserving quality, diversity, and synchronization.
    • Distilled a 40-step teacher into a 4-step student with progressive discrete consistency warm-up, continuous consistency refinement, and joint consistency–distribution matching, achieving 20.1× generator speedup at 512×768 resolution.
    • Combined W8A8 quantization, fused operators, and modality-aware sparse attention to achieve 54.67× generator-only speedup at 1024×1792 resolution on a single NVIDIA H20 (318.74s → 5.83s).
  • From Perception to Planning: Evolving Ego-Centric Task-Oriented Spatiotemporal Reasoning via Curriculum Learning
    Xiaoda Yang*, Yuxiang Liu*, Shenzhou Gao, Can Wang, Jingyang Xue, Lixin Yang, Yao Mu, Tao Jin, Shuicheng Yan, Zhimeng Zhang, Zhou Zhao†
    ICML 2026  [Paper | Code]
    • Proposed EgoTSR, a curriculum-based framework for learning task-oriented spatiotemporal reasoning that evolves from explicit spatial understanding to long-horizon planning.
    • Constructed EgoTSR-Data, a large-scale dataset comprising 46 million samples organized into three stages: CoT supervision, weakly supervised tagging, and long-horizon sequences.
    • Achieved 92.4% accuracy on long-horizon logical reasoning tasks, significantly outperforming existing open-source and closed-source state-of-the-art models.
  • Robust Camera Pose Estimation and 3D Human Reconstruction for Sports Events
    Jing Huang, Hanrong Zhuang, Lin Zhang, Yuxiang Liu, Kun Li†
    Technical Report for FIFA Skeleton Light Challenge 2025  [Paper | Slides]
    • Proposed a method extending the RCR (Robust Crowd Reconstruction) framework to video inputs for sports events.
    • Designed a relative camera pose search algorithm with a fast line projector to achieve robustness and efficiency.
    • Refined 3D HVIP to ensure the consistency of human movement and extracted 3D skeletons from SMPL parameters.

🏅 Honors and Awards

  • 2025.06 Champion, FIFA Skeleton Tracking Challenge (CVPR 2025 Workshop).
  • 2024–2025 academic year National Scholarship (Ministry of Education of China).
  • 2023–2024 academic year National Scholarship (Ministry of Education of China).

📖 Education

  • 2023.09 - Present, Undergraduate in Computer Science and Technology, Tianjin University (TJU).
    • Program: Top-notch Talent Training Plan 2.0 (Top-notch Talent Class).