Introducing ClimaX, the first foundation model for weather and climate. A fast and accurate one-stop AI solution for a range of atmospheric science tasks. Paper: https://t.co/9j10Xlvqzv Blog: https://t.co/D00VPds5hE Thread🧵 #ML #Climate #Weather #FoundationModel

Glad to have contributed a small part to this great research led by @XingyouSong :)

Right before #imo2025, together with colleagues from Mountain View, NYC, Singapore, etc, we all gathered at @GoogleDeepMind headquarter in London for our final push for IMO. I believe that week was when all magic happened! We put all individual recipes (that we figured out

(6/6) 📰 Please check out our paper for additional experiments, ablation studies, and discussions. We published the code and will release the checkpoints soon! Paper: https://t.co/0L94scWvau Code: https://t.co/O0Q9Tx8X41 Work done with my amazing co-first authors and advisor

(5/6) 👀 See It in Action PhysiX generates high-fidelity rollouts that stay accurate over time. While other models blur or drift, PhysiX retains sharpness, physical realism, and temporal consistency—even far into the future.

(4/6) 🏆 The Results? Long-Horizon Accuracy PhysiX doesn't just predict the next frame well; its strength is in long-term rollouts. On shear flow, it reduces error by over 97% compared to the best baseline. Across eight diverse tasks from The Well benchmark, it consistently

(3/6) ⚙️How PhysiX Works: A 3-Step Pipeline 1️⃣ Universal Tokenizer: Encodes any physical field (fluids, solids, etc.) into discrete tokens, like words. 2️⃣ Autoregressive Transformer: Predicts the "next token" to simulate system evolution in space and time. 3️⃣ Refinement Module:

(2/6) 🧪 The Data Bottleneck in Physics Foundation models thrive on massive datasets. However, while there are billions of natural images and videos, even the largest physics simulation datasets have only tens of thousands of examples. PhysiX overcomes this by 1) transferring

🚀 Introducing PhysiX: One of the first large-scale foundation models for physics simulations! PhysiX is a 4.5B parameter model that unifies a wide range of physical systems, from fluid dynamics to reaction-diffusion, outperforming specialized, state-of-the-art models.

Seeing text-to-text regression work for Google’s massive compute cluster (billion $$ problem!) was the final result to convince us we can reward model literally any world feedback. Paper: https://t.co/ojqkfrrDi5 Code: https://t.co/xpNRTxEUjF Just train a simple encoder-decoder

(1/6)Our work Reflect-DiT was accepted to #ICCV2025 ! Reflect-DiT allows the model to reflect on its past generations and textual feedback to self-correct and improve, extending reasoning to text-to-image generation.

Since our launch earlier this year, we are thrilled to witness the growing community around dLLMs. The Mercury tech report from @InceptionAILabs is now on @arxiv with more extensive evaluations: https://t.co/DnDxFvoX0E New model updates dropping later this week!

CVPR-bound! ✈️ I'll be presenting CoLLM on Friday, 6/13 (Morning, #364) and looking for my next challenge as a full-time Scientist/Engineer. If you're hiring or just want to chat about exciting research, find me there! My work: https://t.co/0aMjTwpRSr #CVPR2025 #JobHunt

🎉 Our paper “Representations Shape Weak-to-Strong Generalization” is accepted at #ICML2025! We study weak-to-strong generalization (W2SG)—a core problem in superalignment—and offer new insights into the role of models' internal representations in W2SG. 1/

Announcing the 2025 IJCAI Computers and Thought Award winner ✨Aditya Grover @adityagrover_, @InceptionAILabs @UCLA. Dr. Grover is honored for uniting deep generative models, representation learning & RL to advance scientific reasoning. Congratulations! https://t.co/Z3xESFizpi

I am excited to share my new paper "Streaming Sliced Optimal Transport". I introduce the first algorithm for computing sliced Wasserstein distance from sample streams to make sliced optimal transport even faster and more scalable. (1/10)

📢(1/11)Diffusion LMs are fast and controllable at inference time! But why restrict such benefits for processing text data? We are excited to announce LaViDa, one of the first and fastest large diffusion LM for vision-language understanding!!

Thank you for reading! I unfortunately won’t be at #ICLR2025 due to visa issues 🙁, but please check out our poster on Thursday, Apr 24 at 10AM (Halls 3 & 2B, #309). Please also check out our paper if you are interested in exploring more results and discussions:

Finally, LICO scales favorably with model size: larger backbones consistently yield better performance, making it a promising direction for scaling optimization with LLMs.

We also show that simple language instructions improve the model’s ability to reason in new domains, even when the task is non-linguistic.

Interestingly, we found that the pretrained LLM consistently outperforms a scratch model of the same size. This supports the idea that LLMs can transfer pattern-matching capabilities to entirely new domains.