Pinned Tweet
We show that Gaussian RBMs can generate good images just like other generative models, despite the single-layer architecture. Key innovations:
1) Gibbs-Langevin sampling;
2) modified Contrastive Divergence.
Paper: arxiv.org/abs/2210.10318
Code: github.com/lrjconan/GRBM
1/2
Looks super cool! It is exciting to see such a hybrid world model approach works at scale. Big congrats!
Today we are releasing JST-2, the world's first full-stack physics-native world foundation model.
Many world models learn in pixel/video space. We went back to first principles and built the entire model stack in 3D.
Generate a 3D scene in seconds and expand it in real time. Create characters with motion built in, all through one general architecture.
1
9
850
Renjie Liao retweeted
A Chinese-American hero that inspired my parents' generation to study and stay in America.
Prof. Chen Ning Yang, a world-renowned physicist, Nobel Laureate in Physics, Academician of the Chinese Academy of Sciences, Professor at Tsinghua University, and Honorary Director of the Institute for Advanced Study at Tsinghua University, passed away in Beijing due to illness at the age of 103. His life stands as a timeless chapter in human history—one that shines not only for China but for the global community of thinkers and innovators. His legacy will live on forever.
Renjie Liao retweeted
Interested in how to generate realistic human trajectories using diffusion with just one step🤔?
This is now possible with MoFlow, a one-step Flow Matching method accompanied with Implicit Maximum Likelihood Estimation based distillation.
🚀 Join us at #CVPR2025 in Nashville!
If you’ll be at CVPR, come check out our work LatentHOI on hand–object manipulation!
📍 Poster Session 4, #155
📅 Saturday
Want to generate realistic hand–object manipulations for unseen objects?
Check out our new paper: "LatentHOI: On the Generalizable Hand Object Motion Generation with Latent Hand Diffusion".
Appearing at #CVPR2025 in Nashville!
3
384
🚀Our new work (RetroSynFlow) on retrosynthesis:
1) SOTA discrete flow matching that leverages synthons: products->synthons->reactants
2) effective test-time reward steering
Led by my amazing students @RobinY798, @qi_yan98, with stellar collaborators Guy Wolf & @bose_joey
Excited to announce RetroSynFlow, a discrete flow matching framework for diverse and accurate retrosynthesis!
📄Paper available here: arxiv.org/abs/2506.04439
Joint work w/ @qi_yan98 , Guy Wolf, @bose_joey and @lrjconan .
👇 A quick overview 🧵
Ever wondered how card shuffling can help solve combinatorial problems like sorting, jigsaw puzzles, or TSPs?
Our ICLR 2025 (Oral) paper SymmetricDiffusers introduces a novel discrete diffusion model on finite symmetric groups to solve them in a unified and principled way!
Led by my undergrad intern Nick Zhang (@VectorInst) and my PhD student Donglin Yang @ydlin718 at @ECEUBC.
Huge shoutout to them for driving this forward! 🙌
Code: github.com/DSL-Lab/Symmetric…
Paper: arxiv.org/abs/2410.02942
Interested in LLMs for math reasoning? Check out our new work that builds 647K Math QA pairs and A Live Math Benchmark that is kept up-to-date, led by my brilliant students @sadegh_mahdavi4 @LiJonassen and Kaiwen!
🚀 Training an LLM for math reasoning, but high-quality data is scarce and costly?
Check out our paper: “Leveraging Online Olympiad-Level Math Problems for LLM Training & Contamination-Resistant Evaluation.” 📖✨
TL;DR:
🔹 LLM-powered pipeline for scalable, high-quality data collection 🤖
🔹 647K Math QA pairs 📊
🔹 A Live Math Benchmark that is kept up-to-date⏳
Paper: arxiv.org/abs/2501.14275
Code & Data: github.com/DSL-Lab/aops
Project Website: livemathbench.github.io/
Evaluation Leaderboard: livemathbench.github.io/lead…
@askalphaxiv (for QA): alphaxiv.org/abs/2501.14275
5
372
It means a lot to someone who has worked on Hopfield networks and Boltzmann machines. John and Geoff's foundational work has inspired generations of scientists, and it will keep doing so! I now have a compelling reason to keep teaching and researching these models! Huge Congrats!
BREAKING NEWS
The Royal Swedish Academy of Sciences has decided to award the 2024 #NobelPrize in Physics to John J. Hopfield and Geoffrey E. Hinton “for foundational discoveries and inventions that enable machine learning with artificial neural networks.”
26
1,653
Welcome to submit your work to NeurIPS 2024 workshop on Adaptive Foundation Models!
Deadline: Oct. 4 AOE
adaptive-foundation-models.o…
⌛️We extended the submission deadline for the NeurIPS 2024 Workshop on Adaptive Foundation Models to October 4 AoE!
🏁Submission site: openreview.net/group?id=Neur…
🌐Website: adaptive-foundation-models.o…
Organizers: @mengyer, @PaulVicol, @NailaMurray, @lrjconan, @BeidiChen, @weichiuma
6
1,071
Renjie Liao retweeted
Had fun presenting ES-Single at #ICML2023! ES-Single is a method for low-variance, unbiased gradient estimation in unrolled computation graphs, that has lower variance than PES and outperforms it on a range of tasks.
ICML: icml.cc/virtual/2023/poster/…
Colab: tinyurl.com/es-single-colab
Interested in (spectral) GNNs? Our ICLR work Specformer shows that a simple set-to-set Transformer based spectral filter achieves SOTA on various benchmarks, beating message-passing GNNs (e.g., graph transformers).
Paper: arxiv.org/abs/2303.01028
Code: github.com/DSL-Lab/Specforme…
1/N
Graph Laplacian spectrum contains rich (global) information of the graph, which is hard to be learned by local message passing. Our spectral filter takes the spectrum as input (the set of sin&cos encoded eigenvalues) instead of a single eigenvalue, thus being more expressive. 2/N
1
11
1,221




















