Zhenhua Liu (刘振华)

I am currently a Senior Researcher at Microsoft Research Asia (MSRA). Previously, I worked at AMD and Huawei's Noah's Ark Lab. I received my Ph.D. from Peking University in 2022, advised by Prof. Wen Gao and Prof. Siwei Ma. My current research focuses on Embodied AI.

Feel free to contact me if you are interested in my research or potential collaborations.

Academic Services

Area Chair: ICLR, CVPR, NeurIPS
Conference Reviewer: CVPR, ICCV, ECCV, NeurIPS, ICASSP
Journal Reviewer: TPAMI, IJCV, TIP

Publications

Full list on Google Scholar

BCP learns whether to continue an action chunk or replan from a new observation.

Continue or Replan? Bernoulli-Continuation Policy Learning for Adaptive Horizon Execution

Weichen Xu*, Zhenhua Liu*, Lin Luo, Yaobo Liang, Chengtang Yao, Qingyu Mei, Jian Cao, Xixin Cao, Xing Zhang, Jiaolong Yang, Baining Guo

arXiv · 2026

* Equal contribution.

TaskGround grounds the scene, infers task structure, and compiles executable actions.

TaskGround: Structured Executable Task Inference for Full-Scene Household Reasoning

ZhiYuan Feng, Yu Deng, Ruichuan An, Zhenhua Liu, Qixiu Li, Keming Wu, Zhiying Du, Weijie Wang, Haoxiao Wang, Shuang Chen, Sicheng Xu, Yaobo Liang, Jiaolong Yang, Baining Guo

NeurIPS · 2026

Comparison of weak and strong completers for identifying reliable reasoning-step labels.

Athena: Enhancing Multimodal Reasoning with Data-efficient Process Reward Models

Shuai Wang, Zhenhua Liu, Jiaheng Wei, Xuanwu Yin, Dong Li, Emad Barsoum

TMLR · 2026

Jakiro uses decoupled experts to produce diverse draft candidates for speculative decoding.

Jakiro: Boosting Speculative Decoding via Decoupled MoE

Haiduo Huang, Fuwei Yang, Zhenhua Liu, Pengju Ren

ACL · 2026

Overview of adaptive bit switching using Double Rounding and adaptive learning-rate scaling.

Double Rounding: Nearly Lossless Adaptive Bit Switching for QAT

Haiduo Huang, Zhenhua Liu, Tian Xia, Pengju Ren

AAAI · 2026

SpecVLM compresses visual tokens for an efficient draft model and verifies predictions with the target model.

SpecVLM: Fast Speculative Decoding in Vision-Language Models

Haiduo Huang, Fuwei Yang, Zhenhua Liu, Xuanwu Yin, Dong Li, Pengju Ren, Emad Barsoum

arXiv · 2025

Kangaroo uses a shallow sub-network and early exits during speculative drafting.

Kangaroo: Lossless Self-Speculative Decoding for Accelerating LLMs via Double Early Exiting

Fangcheng Liu, Yehui Tang, Zhenhua Liu, Yunsheng Ni, Duyu Tang, Kai Han, Yunhe Wang

NeurIPS · 2024

Comparison of GhostNetV2 and GhostNetV3 architectures and re-parameterized training blocks.

GhostNetV3: Exploring the Training Strategies for Compact Models

Zhenhua Liu, Zhiwei Hao, Kai Han, Yehui Tang, Yunhe Wang

arXiv · 2024

A timeline of milestones in the development of Transformer models.

A Survey on Vision Transformer

Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chunjing Xu, Yixing Xu, Zhaohui Yang, Yiman Zhang, Dacheng Tao

IEEE TPAMI, 45(1): 87–110 · 2023