A sunflower wearing sunglasses — Ligeng Zhu's profile photo

Ligeng Zhu

I am a Researcher at NVIDIA, fortunately working with Prof. Song Han. Before NVIDIA, I was at MIT (freezing Boston) with Prof. Song Han, Simon Fraser University (warmful Vancouver) with Prof. Brian Funt and Prof. Ping Tan, and Zhejiang University (beautiful Hangzhou).

I am currently leading Humanize (Flow Harness Framework) and KDA (Kernel Design Agents) at NVIDIA Research. My research focuses on scalable and efficient agents, and my mission is five "one"s:

100T Model + 1K TPS + 1M Agents + 1M Dockers + 1M Mins.

If you find any research interests that we might share, feel free to drop me an email. I am always open to potential collaborations.

.- -.. / .- ... - .-. .- .-.-.- / .-.. . - .----. ... / ... -.-. .- .-.. . -.-.--

Latest update on

News

Full News
  • Completed full Lean formalization of the Hamilton–Perelman proof of Thurston’s Geometrization conjecture (beyond Poincaré) with the Humanfia team and collaborators (Geometry library, Humanize, Post).
  • Regained first place on the LeanEval leaderboard with updated Humanize workflows after ICLR (Leaderboard, Post).
  • KDA optimized Kimi Delta Attention kernels up to 2.96× over FlashKDA (with Humanize2 / Flame Chase); also helped Databricks reach #1 across SolExecBench tracks in about one week (Details, KDA).
  • Reached 100% (672/672) on PutnamBench and took #1 on LeanEval using Humanize2.
  • Powered by Humanize + Kimi-K3, achieved a perfect score (6/6, 42/42) on IMO 2026 Lean formal proofs in one run at lower cost than GPT-5.6 (Results, Post).
  • Presented about token maxing with Humanize at NVIDIA, invited by SGLang (Slides, Talks).
  • KDA (Kernel Design Agents) powered HAN Lab Kernel Mafia to #1–#3 rankings (MoE, DSA, GDN) in the Pure Agent track of the MLSys FlashInfer Kernel Contest, with minimal prior CUDA experience (Post, KDA).
  • Our PockEngine is repoted by MIT News!

Selected Publications

Full Publications

A Lean Formalization of the Hamilton--Perelman Proof of Thurston's Geometrization Conjecture

Yuan Liao, Ziyang Qin, Jui-Hui Chung, Ayush Khaitan, Zhengyang Zhang, Hongzhou Lin, Zijian Zhang, Ligeng Zhu, Peng Lu, Bennett Chow

arXiv preprint, 2026

Humanize Olympiad Agents: A Multi-Session Workflow for Olympiad Problem Solving with Solve, Review, Wiki, Verify

Zhengyang Zhang, Jing Xiong, Hongzhou Lin, Juihui Chung, Zijian Zhang, Song Han, Sihao Liu, Ligeng Zhu

arXiv preprint, 2026

Humanize MLE Agents: Multi-Agent Session Alternation Beyond Single-Agent Limits

Changye Li, Zijian Zhang, Muheng Li, Yiduo Pang, Tingjun Wu, Menghan Li, Yitong Liu, Hongkai Guo, Mengchi Chen, Chenyu He, Kaisheng Ma, Wei Xu, Song Han, Sihao Liu, Ligeng Zhu

arXiv preprint, 2026

Humanize: Judgement Engineering for Agentic Coding

Sihao Liu, Ligeng Zhu, Zijian Zhang, Dongyun Zou, Zhengyang Zhang, Changye Li, Song Bian, Song Han, Tony Nowatzki

arXiv preprint, 2026

NVILA: Efficient Frontier Visual Language Models

Zhijian Liu, Ligeng Zhu, Baifeng Shi, Zhuoyang Zhang, Yuming Lou, Shang Yang, Haocheng Xi, Shiyi Cao, Yuxian Gu, Dacheng Li, Xiuyu Li, Haotian Tang, Yunhao Fang, Yukang Chen, Cheng-Yu Hsieh, De-An Huang, An-Chieh Cheng, Jinyi Hu, Sifei Liu, Ranjay Krishna, Pavlo Molchanov, Jan Kautz, Hongxu Yin, Song Han, Yao Lu

IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

On-Device Training Under 256KB Memory

Ji Lin*, Ligeng Zhu*, Wei-Ming Chen, Wei-Chen Wang, Chuang Gan, Song Han

(* denotes equal contribution)

Advances in Neural Information Processing Systems (NeurIPS), 2022

Media coverage:

Deep Leakage from Gradients

Ligeng Zhu, Zhijian Liu, Song Han

Advances in Neural Information Processing Systems (NeurIPS), 2019

Media coverage:

ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardware

Han Cai, Ligeng Zhu, Song Han

International Conference on Learning Representations (ICLR), 2019

Media coverage:

Students Collaborated with

  • Zhengyang ZhangYao Class Undergraduate at Tsinghua University → MIT Visiting
  • Changye LiUndergrad at Peking University → PhD at Tsinghua University
  • Zheng DuPhD at GaTech
  • Zijian ZhangUndergrad at Tsinghua University → PhD at Tsinghua University
  • Yunhao FangMaster at UCSD → ByteDance Seed → Physical Intelligence
  • Guanqi ZhanPhD at VGG Oxford → NVIDIA Research
  • Jiashu HanUndergrad at Columbia University → MIT EECS PhD
  • Lanxiang HuMaster at Columbia University → PhD at UCSD ECE
  • Yaoyao DingUndergrad at Shanghai Jiaotong University → PhD at University of Toronto

Talks & Presentations

  • Presented about token maxing with Humanize at NVIDIA, invited by SGLang (Slides, Talks).

  • AutoML for Efficient Neural Architecture Design (Slides)
    @ OpenPower Summit, Polarr Tech

  • Scalable and Secure Machine Learning for Edge Devices @ Qualcomm

  • Neural Architecture Designs @ UIUC IFP Group (Slides)

  • Proxylessly Specialize CNN for Hardware @ IBM-MIT Watson Events (Poster)

  • Sparsely Aggregated Convolutional Networks (Slides)
    @ UBC Vision Group, Deephi Tech

  • Invited lectures about deep learning (Lecture1, Lecture2)
    @ SFU Computer Vision Course (CMPT-412), ZJU Programming Group

Open-source Projects (Selected)

My life (both academic and daily) is greatly powered by open source projects. To thank their selfless effort, I embrace open source as much as possible. Please refer to my github for a complete list of projects.

VILA

Stars: 3.9k

Vision-language models for multimodal AI on edge devices, servers, and the cloud.

SANA

Stars: 9.3k

Efficient high-resolution image generation with linear diffusion transformers.

PyTorch-OpCounter

Stars: 5,082

Count the MACs / FLOPs of your PyTorch model.
Avaliable through PyPi: pip install thop

Services

Review papers for:

2026
CVPR, ICLR, NeurIPS, ECCV, ACL
2025
CVPR, ICLR, ICCV
2024
CVPR, ICLR, ICML
2023
NeurIPS, CVPR, ICLR, ICML
2022
NeurIPS, CVPR
2021
NeurIPS, ICCV, ICML, ACL
2020
NeurIPS, CVPR, AAAI
2019
NeurIPS, ICCV, CVPR
Journals
T-PAMI, IEEE Micro