🚀 Excited to share our new paper!
SiamJEPA: On the Role of Siamese Student Encoders in JEPA
https://t.co/IhK1egFEOL
JEPA has emerged as a promising framework for self-supervised representation learning.
In this work, we ask a simple but important question:
What is the role of Siamese student encoders in JEPA?
To answer this question, we propose SiamJEPA, a JEPA framework with Siamese student encoders and an EMA teacher.
A short thread 👇
Here's a beautiful and clean work led by Xianliang and Zihan. I love this work because it demonstrates that momentum is not a "decoration" to Muon but an essential component. Take a look at the paper as well as the landing page!
🧠Why does Muon do momentum before orthogonalization?
✨Our key insight: momentum acts as a spectral filter for the matrix-valued gradient, yielding a more reliable update for the orthogonalization step.
📝Paper: https://t.co/mqjxCrcSIV
🌐Project: https://t.co/Cep6HK0oAA
🧠Why does Muon do momentum before orthogonalization?
✨Our key insight: momentum acts as a spectral filter for the matrix-valued gradient, yielding a more reliable update for the orthogonalization step.
📝Paper: https://t.co/mqjxCrcSIV
🌐Project: https://t.co/Cep6HK0oAA
Our #ICML2026 paper with @levelfour_ and @nanjing_yzcao gives a “small-surrogate-loss + path-length” guarantee for non-stationary online structured prediction! The latest version is now available on arXiv:
https://t.co/z7tL6feeCS
I’m excited to share a new paper on online inverse optimization/contextual recommendation! 🚀
It simplifies our NeurIPS 2025 method, keeps the best-known O(d log T) regret, removes costly projections via a second-order perceptron idea, and is robust to suboptimal-action feedback.
A perspective on flow matching that particularly resonates with me is a statistical-physics-flavored one:
the proximal formulation makes the denoising dynamics look like a relaxation process in an effective potential landscape.
Check out our recent work: https://t.co/8RpHnc6fHs
After nearly 3 years, I completed my COLT2023 paper as a journal paper, which appeared yesterday in JMLR🙌 This was quite a longer journey than expected initially and finally...
Proper losses in multiclass case, regret bounds, downstream tasks, and more!
https://t.co/LuOJzO4KYi
Recent updates 🙌
💬 Our paper on OT x NLP will appear in #ACL2023 (see https://t.co/h9pRqgcaHL)
🧑💻 My solo paper on proper loss x convexity will appear in #COLT2023 (see pic)
Now's the time I need an avatar to be in Toronto 🇨🇦 and Bangalore 🇮🇳 at the same time 🔥🔥