Efficient AI Researcher

Yuexiao Ma (马跃萧)

Research Fellow at MMLab@NTU, S-Lab · Ph.D. at Xiamen University

I make large generative and vision models faster, smaller, and more practical. My work spans low-bit quantization, model compression, and efficient video generation. I always welcome research collaborations and academic exchanges. Please feel free to email me with any questions.

Google Scholar GitHub Email

Yuexiao Ma smiling outdoors at a city viewpoint

About

I am a Research Fellow at MMLab@NTU, S-Lab, working with Prof. Chen Change Loy. I received my Ph.D. in Intelligent Science and Technology from Xiamen University, advised by Prof. Rongrong Ji and Assoc. Prof. Xiawu Zheng.

Previously, I spent more than three years with ByteDance's quantization team, translating compression research into low-cost, low-latency inference for generative and foundation models.

Research

Making intelligence efficient by design.

I study the algorithms and systems that remove unnecessary computation without sacrificing model capability.

Low-bit Quantization

Post-training and mixed-precision methods for LLMs, vision transformers, diffusion models, and production inference.

Efficient Video Generation

Temporal and flow-aware caching strategies that accelerate autoregressive generation while preserving visual quality.

Model Compression

Theory-guided pruning, outlier handling, and hardware-aware optimization for deployable neural networks.

Publications

^* Equal contribution · ^† Corresponding author

Google Scholar

ICLR 2026Video generation

Flow Caching for Autoregressive Video Generation

Yuexiao Ma^*, Xuzhe Zheng^*, Jing Xu^*, Xiwei Xu, Feng Ling, Xiawu Zheng, Huafeng Kuang, Huixia Li, Xing Wang, Xuefeng Xiao, Fei Chao, Rongrong Ji^†

Flow-aware feature reuse accelerates autoregressive video generation while preserving temporal quality.

Paper Code

MotionCache frame importance map visualization

ICML 2026Video generation

Motion-Aware Caching for Efficient Autoregressive Video Generation

Jing Xu^*, Yuexiao Ma^*, Xuzhe Zheng, Xing Wang, Shiwei Liu, Chenqian Yan, Xiawu Zheng, Rongrong Ji, Fei Chao, Songwei Liu^†

A coarse-to-fine cache policy uses inter-frame motion to allocate computation where it matters.

Paper Code

Overview of training-free multimodal orchestration

ICML 2026Multimodal LLMs

Training-Free Multimodal Large Language Model Orchestration

Tianyu Xie, Yuexiao Ma, Yuhang Wu, Wang Chen, Jiayi Ji, Tat-Seng Chua, Xiawu Zheng^†, Rongrong Ji

A controller composes specialized modality experts into a unified training-free multimodal system.

Paper Code

ICML 2026Reasoning benchmark

A²RBench: An Automatic Paradigm for Formally Verifiable Abstract Reasoning Benchmark Generation

Qingchuan Ma, Yuexiao Ma, Yongkang Xie, Tianyu Xie, Xiawu Zheng^†, Rongrong Ji

An automated pipeline generates abstract reasoning tasks with programmatic verification.

Paper Code

Overview of polybasic speculative decoding

ICML 2025Speculative decoding

Polybasic Speculative Decoding Through a Theoretical Perspective

Ruilin Wang, Huixia Li, Yuexiao Ma, Xiawu Zheng, Fei Chao, Xuefeng Xiao, Rongrong Ji^†

A theoretical framework generalizes draft-and-verify decoding to multiple interconnected draft models.

Paper

IJCV 2025Network pruning

An Information Theory-inspired Strategy for Automatic Network Pruning

Xiawu Zheng^*, Yuexiao Ma^*, Teng Xi, Gang Zhang, Errui Ding, Yuchao Li, Jie Chen, Yonghong Tian, Rongrong Ji^†

Information-theoretic layer importance enables one-pass, search-free automatic pruning.

Paper Code

Relationship between reconstruction granularity, quantization loss, and model performance

ICML 2024Vision quantization

Outlier-aware Slicing for Post-Training Quantization in Vision Transformer

Yuexiao Ma, Huixia Li, Xiawu Zheng, Feng Ling, Xuefeng Xiao, Rui Wang, Shilei Wen, Fei Chao, Rongrong Ji^†

Outlier-aware channel slicing substantially improves accurate INT4 deployment of vision transformers.

Paper

ICLR 2024LLM quantization

AffineQuant: Affine Transformation Quantization for Large Language Models

Yuexiao Ma, Huixia Li, Xiawu Zheng, Feng Ling, Xuefeng Xiao, Rui Wang, Shilei Wen, Fei Chao, Rongrong Ji^†

Learned affine transformations reduce activation outliers and post-training quantization error.

Paper Code

CVPR 2023Quantization theory

Solving Oscillation Problem in Post-Training Quantization Through a Theoretical Perspective

Yuexiao Ma, Huixia Li, Xiawu Zheng, Xuefeng Xiao, Rui Wang, Shilei Wen, Xin Pan, Fei Chao, Rongrong Ji^†

A theoretical diagnosis of PTQ oscillation leads to mixed reconstruction granularity optimization.

Paper Code

AAAI 2023Mixed precision

OMPQ: Orthogonal Mixed Precision Quantization

Yuexiao Ma, Taisong Jin^†, Xiawu Zheng, Yan Wang, Huixia Li, Yongjian Wu, Yunsheng Wu, Guannan Jiang, Wei Zhang, Rongrong Ji

Orthogonality-guided sensitivity estimates allocate mixed precision without iterative search.

Paper Code

BVP 2019Applied mathematics

Positive Solutions of Fourth-order Problems with Dependence on All Derivatives in Nonlinearity under Stieltjes Integral Boundary Conditions

Yuexiao Ma, Chenyang Yin, Guowei Zhang^†

Existence results for positive solutions of fourth-order boundary value problems under Stieltjes conditions.

Paper

Experience & education

Present

Research

Research Fellow

MMLab@NTU · S-Lab

Working with Prof. Chen Change Loy

2022 — 2025

Industry

Research Intern · Quantization Team

ByteDance Data · Beijing

2022 — 2026

Education

Ph.D. · Intelligent Science and Technology

Xiamen University · Department of Artificial Intelligence

Advisors: Prof. Rongrong Ji and Assoc. Prof. Xiawu Zheng

2020 — 2022

Education

M.S. · Intelligent Science and Technology

Xiamen University · Department of Computer Science

2016 — 2020

Education

B.S. · Mathematics and Applied Mathematics

Northeastern University · Shenyang

Academic service

Journal reviewer

IEEE TIP · IEEE TMM

Conference reviewer

ECCV · NeurIPS · ICML · ICLR · AAAI · CVPR

Recognition

Top Reviewer

NeurIPS 2025 · ICLR 2026 · ICML 2026

Research · Collaboration · Ideas

Let's make ambitious models practical.

bobma@stu.xmu.edu.cn

Yuexiao Ma (马跃萧)

Making intelligence efficient by design.

Low-bit Quantization

Efficient Video Generation

Model Compression

Publications

Flow Caching for Autoregressive Video Generation

Motion-Aware Caching for Efficient Autoregressive Video Generation

Training-Free Multimodal Large Language Model Orchestration

A2RBench: An Automatic Paradigm for Formally Verifiable Abstract Reasoning Benchmark Generation

Polybasic Speculative Decoding Through a Theoretical Perspective

An Information Theory-inspired Strategy for Automatic Network Pruning

Outlier-aware Slicing for Post-Training Quantization in Vision Transformer

AffineQuant: Affine Transformation Quantization for Large Language Models

Solving Oscillation Problem in Post-Training Quantization Through a Theoretical Perspective

OMPQ: Orthogonal Mixed Precision Quantization

Positive Solutions of Fourth-order Problems with Dependence on All Derivatives in Nonlinearity under Stieltjes Integral Boundary Conditions

Research Fellow

Research Intern · Quantization Team

Ph.D. · Intelligent Science and Technology

M.S. · Intelligent Science and Technology

B.S. · Mathematics and Applied Mathematics

IEEE TIP · IEEE TMM

ECCV · NeurIPS · ICML · ICLR · AAAI · CVPR

Top Reviewer

Let's make ambitious models practical.

A²RBench: An Automatic Paradigm for Formally Verifiable Abstract Reasoning Benchmark Generation