Junchen Liu

@JunchenLiu77

Research Scientist Intern @NVIDIA | PhD student @UofT @VectorInst.

Toronto

junchenliu77.github.io

Joined September 2022

Pinned
Junchen Liu
@JunchenLiu77
Feb 25
Continual learning and online adaptation are often framed as the next frontier of AI. 🚀 Modern architectures use Test-Time Training (TTT) to memorize key-value pairs on the fly via gradient descent, or so we thought. To test this memorization hypothesis, we replaced gradient
Junchen Liu
@JunchenLiu77
Jul 6
Very well deserved work! Congrats to @cindy_x_wu @JunGao33210520 @jonLorraine9 and the team!!
Jonathan Lorraine
@jonLorraine9
Jul 6
Replying to @icmlconf
So honored that MOTIVE was named an ICML 2026 Outstanding Paper Honorable Mention! 🎉 Motion Attribution for Video Generation 👇 research.nvidia.com/labs/sil/proje… TL;DR: We propose Motive, a scalable, motion-centric data attribution framework for video generation to identify which
Junchen Liu
@JunchenLiu77
Jul 3
I’ll be at #ICML2026 next week presenting this work. icml.cc/virtual/2026/p… See you in Seoul! 🇰🇷
Junchen Liu
@JunchenLiu77
Feb 25
Continual learning and online adaptation are often framed as the next frontier of AI. 🚀 Modern architectures use Test-Time Training (TTT) to memorize key-value pairs on the fly via gradient descent, or so we thought. To test this memorization hypothesis, we replaced gradient
Junchen Liu
@JunchenLiu77
Jun 7
My favorite paper so far this year: a simple idea, solid experiments, and overwhelmingly strong results.
Jiawei Yang
@JiaweiYang118
May 1
Two months ago, I vaguely posted a number: 0.9 FID, one-step, pixel space. Now it is 0.75, and can be even lower. Many wonder how. I thought it might end as a small FID prank: simple and deliberate. It started with one question: can FID be optimized directly, and what does it
Junchen Liu
@JunchenLiu77
Jun 3
Excited to share FlashDreams! It pushes world models closer to real-time interaction with faster AR inference, streaming generation, multi-GPU serving, and low-latency deployment. World models need strong runtime infrastructure, not just better generators, and FlashDreams is
Ruilong Li
@ruilong_li
Jun 3
World models are moving beyond offline generation towards interactive, real-time experiences. Introducing ⚡FlashDreams⚡: an open-source high-performance inference and serving library built for autoregressive world models: 🔥 Up to 3.10× faster LingBot-World inference 🔥 Up to
00:00