About Me

I am Chunyu Liu (刘春宇), a third-year undergraduate student in Yao Class at Tsinghua University, advised by Alex Lamb. My research lies at the intersection of language modeling, ML efficiency, and self-evolving agents. For collaborations or discussions, feel free to reach me via email.

Outside the lab, my artistic side lives under the name 瑾瑜當年. 觀瀾 is a quiet archive for my essays and poems. You can get to know me better through my Music and Paintings.

Education

B.S. in Computer Science and Technology

IIIS (Yao Class), Tsinghua University

2024-2028 (expected)

GPA 3.9 / 4.0

• A+: Advanced Computer Graphics, Introduction to Large Language Model Applications, Type-safe Modern System Practice, Introduction to Programming in C/C++.
• A: Deep Learning, Machine Learning, Natural Language Processing, Embodied Artificial Intelligence, AI+X Computing Acceleration, Introduction to Computer Systems, 3 more courses.

Publications

* Equal contribution.

COLM

VoidPadding: Let [VOID] Handle Padding in Masked Diffusion Language Models so that [EOS] Can Focus on Semantic Termination

Chunyu Liu, Zhengyang Fan, Kaisen Yang, Alex Lamb

COLM 2026 Workshop on Efficient Reasoning (Spotlight), 2026

PDF Code

ArXiv

JustQuant: You Don't Need Smoothing, SVD, or Rotation for 4-Bit Activation Quantization

Kaicheng Yang*, Kaisen Yang*, Chunyu Liu*, Xianglong Yan, Haotong Qin, 6 more authors

ArXiv preprint, 2026

PDF Code

ArXiv

Frozen in a Frame: The Velocity Blind Spot in JEPA World Models

Tinghe Zhang, Chunyu Liu, Yu Leon Liu, Zerui Zhao, Jiaheng Chen, 4 more authors

ArXiv preprint, 2026

PDF Code

COLM

Understanding Is Done Early: A Depth Division of Labor in Large Language Models and Its Use for Unbounded-Context Memory

Hanzuo Liu, Xuan Qi, Chunyu Liu, Haotian Zhong, Yulong Wang, 4 more authors

COLM 2026 Workshop on Efficient Reasoning, 2026

PDF Code

ArXiv

A Survey of Efficient Attention Methods: Hardware-Efficient, Sparse, Compact, and Linear Attention

Jintao Zhang, Rundong Su*, Chunyu Liu*, Jia Wei*, Ziteng Wang*, 26 more authors

ArXiv preprint, 2025

PDF

Research Experience

Tsinghua College AI

Research Intern: 2026.3 - Present.

Topics: Diffusion Language Models and Agent Memory.

Research Advisor: Prof. Alex Lamb.

Tsinghua SAIL Group

Research Intern: 2025.6 - 2026.2.

Topics: Sparse Attention, Linear Attention, and Quantization.

Research Advisor: Prof. Jianfei Chen and Prof. Jun Zhu.

Industry Experience

Einsia

AI Research Engineer: 2026.8 - Present.

Topics: Physical Agents, Benchmarking, and AgentGit R&D.

Ubiquant IQuest Lab

Pretraining Algorithms Intern: 2026.6 - 2026.9.

Topics: Sparse Attention.

Selected Projects

AgentGit

An agent-native development workflow for repository search, collaboration, and research-and-development work.

Website

Moe Traits 2 Companion

A multimodal AI desktop companion combining screen and camera perception, tool-calling agents, persistent memory, and a character-creation workbench.

Code

HACT

A hierarchical ACT policy for long-horizon, language-conditioned manipulation, reaching 87% simulated success.

Code

Awards

• Silver Medal, 21st Chinese Girls’ Mathematical Olympiad · 2022
• Freshman Scholarship, Tsinghua University · 2024
• Artistic Excellence Scholarship, Tsinghua University · 2025
• Xuetang Scholarship, Tsinghua University · 2024–Present