🚀 I'm excited to announce the Stanford Data-Centric AI Workshop 2026 (aka the Stanford Graph Learning Workshop), happening Tuesday, November 3, 2026 at @Stanford University (with online livestream). Free registration! Submit a poster. 📍
This year's workshop will spotlight three fast-moving frontiers in data-centric AI:
-- AI Foundations — Making frontier models fast and scalable
-- AI for Science and Biology — Agents and foundation models across the scales of biology
-- Relational and Tabular Foundation Models — Learning directly on the structured data that runs the world
We're bringing together Stanford researchers and industry partners for a full day of research talks, case studies, and collaborative discussion. Whether you're working in industry, academia, or startup land, there will be something to spark your curiosity and drive your work forward.
🔍 Want to share your work? Our Call for Posters and Demos is also open now.
✅ Register now (free): snap.stanford.edu/graphlearn…
📅 Save the date: Nov 3, 2026
I'm hiring a Research Scientist at @StanfordAILab !
Lead research, write and manage grants, teach, and mentor PhD students in my lab. A real growth opportunity, working directly with me.
3-year role, in person, $125K–$225K base.
Apply: careersearch.stanford.edu/hc…
We’re hiring postdocs at @StanfordAILab! Join the Leskovec Lab, in collaboration with the Good Lab, to build foundation models and AI agents for biology.
Help turn multimodal biomedical data into scientific discoveries.
Apply: snap.stanford.edu/apply/inde…
🦍 Memorilla treats memory as a modality for LLMs. Memorilla turns everything a frozen LLM needs to remember into just 16 memory tokens. It uses up orders of magnitude fewer tokens while being faster and more accurate than full context inference. 🎯 Pushes the Pareto frontier over RAG and agentic memory in personalization, long-document and knowledge QA.
🦍 Introducing Memorilla: memory as a modality for LLMs
Memorilla turns everything a frozen LLM needs to remember into just 16 memory embeddings
⚡ It uses up to 4,212× fewer tokens and is 27× faster than full context inference while more accurate
🎯 Pushes the Pareto frontier over RAG and agentic memory in personalization, long-document and knowledge QA
A thread 🧵:
🚀 It's time! The Stanford Data-Centric AI Workshop 2026 is kicking off NOW!
Join us for a full day of research talks, case studies, and community connection — all centered around this year's themes:
⚡ AI Foundations – making frontier models fast and scalable
🧬 AI for Science and Biology – agents and foundation models across the scales of biology
🔗 Relational and Tabular Foundation Models – learning directly on the structured data that runs the world
🎥 Tune in live: youtube.com/live/xQJEsCH0kIA
📍 Stanford University | 🕗 Nov 3, 2026 | 8:30 AM – 6:00 PM PT
📢 The Full Schedule for the Stanford Data-Centric AI Workshop 2026 is Here!
We have an exciting lineup of speakers from industry and academia, including:
🔹 @SAP – semantics-aware foundation models for enterprise data
🔹@NVIDIA – tabular foundation models
🔹 Prof. @tatsu_hashimoto (@Stanford) – data-efficient language modeling
🔹 Prof. @james_y_zou (@Stanford) – learning a physiology foundation model
They are joined by Stanford researchers presenting new work across AI Foundations, AI for Science and Biology, and Relational and Tabular Foundation Models, plus a poster slam and poster sessions over lunch and at the evening happy hour.
📅 Tuesday, November 3, 2026 | 8:30 AM – 6:00 PM PT
📍 Stanford University + Online
🔗 Full schedule & free registration: snap.stanford.edu/graphlearn…
See you next week!
Introducing 💥RT-J💥, a System 1 model for Data
Science: put raw data in context, predict in one forward pass⚡No feature engineering, no model training, just accurate predictions.
Excited to release models, pretraining data+code and NeurIPS paper under the Stanford Tabular and Relational (STAR) ecosystem (links in comments).
How did we build this? 👇🧵
𝗖𝗮𝗻 𝗮 𝗺𝘂𝗹𝘁𝗶𝗺𝗼𝗱𝗮𝗹 𝗺𝗼𝗱𝗲𝗹 𝗯𝗲𝗰𝗼𝗺𝗲 𝗮 𝗯𝗲𝘁𝘁𝗲𝗿 𝗶𝗺𝗮𝗴𝗲 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗼𝗿 𝗯𝘆 𝗹𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝗳𝗿𝗼𝗺 𝗶𝘁𝘀 𝗼𝘄𝗻 𝗳𝗲𝗲𝗱𝗯𝗮𝗰𝗸? 🚀
Excited to share 𝗨𝗻𝗶𝗘𝘃𝗼-𝗩𝗟, our new work on on-policy self-distillation for multimodal model self-improvement!
The motivation is simple: as generation and understanding come together in multimodal systems, models can provide feedback on their own outputs. Can that feedback become a useful training signal—turning their ability to evaluate into improvements in generation?
📈 𝗧𝗵𝗲 𝗿𝗲𝘀𝘂𝗹𝘁𝘀
Building on Qwen-image-2512, UniEvo-VL improves:
• GenEval: 𝟬.𝟳𝟰𝟳 → 𝟬.𝟴𝟬𝟴
• GenEval2 Soft-TIFA: 𝟯𝟮.𝟵𝟳 → 𝟯𝟱.𝟱𝟯
These gains show the potential of learning from self-generated feedback. But the boundaries are just as interesting.
🔍 𝗪𝗵𝗮𝘁 𝘀𝗲𝘁𝘀 𝘁𝗵𝗲 𝗰𝗲𝗶𝗹𝗶𝗻𝗴?
Experiments with stronger external critics suggest room for further improvement.
This raises a central challenge: producing feedback is only the beginning. That feedback must be reliable enough to guide learning. Understanding what makes a good critic matters for building a better generator.
🔍 𝗗𝗼𝗲𝘀 𝗶𝗺𝗽𝗿𝗼𝘃𝗲𝗺𝗲𝗻𝘁 𝘁𝗿𝗮𝗻𝘀𝗳𝗲𝗿 𝗮𝗰𝗿𝗼𝘀𝘀 𝘁𝗮𝘀𝗸𝘀?
Not uniformly. Text-rendering results remain mixed, even as the overall generation benchmarks improve.
Progress on one benchmark does not mean every capability improves alongside it—a useful boundary to understand when studying self-improving systems.
For me, the broader research question is:
𝗪𝗵𝗲𝗻 𝗶𝘀 𝗮 𝗺𝗼𝗱𝗲𝗹’𝘀 𝗼𝘄𝗻 𝗳𝗲𝗲𝗱𝗯𝗮𝗰𝗸 𝗿𝗲𝗹𝗶𝗮𝗯𝗹𝗲 𝗲𝗻𝗼𝘂𝗴𝗵 𝘁𝗼 𝗵𝗲𝗹𝗽 𝗶𝘁 𝗶𝗺𝗽𝗿𝗼𝘃𝗲?
UniEvo-VL takes a step toward answering this question for multimodal generation, exploring both the opportunities and the limitations.
Grateful to all my collaborators for making this work possible! If you’re working on self-distillation, multimodal post-training, or self-improving models, I’d love to hear your perspective.
📄 Paper: arxiv.org/abs/2609.38721
🤗 Hugging Face: huggingface.co/papers/2609.3…
If you find the work interesting, a repost or an upvote on Hugging Face would help more researchers discover it. Thanks for your support! 🙏
#UniEvoVL#SelfDistillation#MultimodalAI#ImageGeneration#PostTraining
Not that long ago, building a machine learning system meant carefully designing features, choosing a model architecture, training it from scratch, tuning hyperparameters, and repeating this process for every new problem. Then foundation models changed how we think about text, images, and increasingly other modalities: pretrain once, then adapt to new tasks through context.
The same transition is now happening for structured data.
With Kumo Tabular, you provide a table with labeled examples and rows you want predictions for. The model produces predictions in a single forward pass --— with no task-specific training, no fine-tuning, and no feature engineering.
What makes this especially exciting to me is that this is not just a new model, but part of a rapidly growing research ecosystem around tabular foundation models. There is tremendous innovation happening across academia and industry in architectures, synthetic pretraining, in-context learning, evaluation, and efficient inference.
NVIDIA wants to be an active part of that ecosystem —-- contributing research, releasing models openly, and building infrastructure that helps the community push the field forward.
There are several aspects of the work I find particularly interesting.
** The models are pretrained entirely on synthetic tables generated from structural causal models, allowing us to expose them to enormous diversity without training on customer data or benchmark datasets. The largest model sees more than 100 million synthetic tables during pretraining.
** The resulting models are both accurate and efficient. Across major tabular benchmarks, Kumo Tabular improves upon strong existing approaches while requiring no per-dataset training. On TabArena, for example, Kumo Tabular Large sits on the accuracy–speed Pareto frontier and is 17× faster at prediction than LimiX-2.
To me, the bigger story is the direction ML is moving: from hand-built models for individual tasks to pretrained models that learn broad representations of a domain and can solve new problems from context. We have seen this transformation in language and vision. It is exciting to see it now reaching the enormous world of structured data.
Huge congratulations to the team — and to the broader tabular foundation model research community whose ideas and work are making this new paradigm possible.
We’re excited to contribute, learn, and help build this ecosystem together.
I’m very excited to share @nvidia Kumo Tabular, a new family of foundation models for tabular data.
Kumo Tabular establishes the new Pareto frontier across the entire accuracy–inference-time tradeoff.
Just as importantly, we are releasing it openly: open weights, open-source software, and a permissive license for commercial use.
HuggingFace: huggingface.co/nvidia/Kumo-T…
GitHub: github.com/NVIDIA/structured…
Introducing CUDA Rust!
CUDA Rust lets you write GPU kernels natively in Rust, not just launch them from it.
Two paths: cuda-oxide for SIMT kernels compiled to PTX, and cutile-rs for Tile-based programming on stable Rust. Both can catch aliasing errors at compile time.
Technical blog: nvda.ws/4hm1bHS
Excited to see this out! Relational learning needs strong benchmarks, standardized evaluation, and easy ways to bring methods to real-world relational data.
We’re happy to announce our first release in relational learning at Prior Labs, continuing our commitment to open science.
We open-source three pieces of software that we expect to accelerate research in the field towards meaningful, real-world impact.
First and foremost, we release 𝗥𝗲𝗹𝗔𝗿𝗲𝗻𝗮-α: a unified framework for running and comparing baselines on RelBench v1 tasks. Based on learnings from tabular benchmarks like TabArena, we are standardizing data loading, evaluation protocols, tuning regimes, and adding support for systems with custom tuning.
We also open-source 𝗧𝗮𝗯𝗣𝗙𝗡-𝗥𝗲𝗹: our relational harness for TabPFN-3. We initialize the (living) RelArena-α leaderboard with TabPFN-Rel and a comprehensive set of baselines. The rankings at the time of release are:
• 𝗧𝗮𝗯𝗣𝗙𝗡-𝗥𝗲𝗹 is the No. 1 model submission
• 𝗥𝗧-𝗣𝗹𝘂𝗥𝗲𝗹 is the No. 1 system submission
Last but not least, we open-source an alpha version of the 𝗥𝗲𝗹𝗮𝘁𝗶𝗼𝗻𝗮𝗹 𝗣𝗿𝗲𝗱𝗶𝗰𝘁𝗶𝘃𝗲 𝗜𝗻𝘁𝗲𝗿𝗳𝗮𝗰𝗲 (𝗥𝗣𝗜): enabling you to easily specify prediction tasks on your own relational database and then run any RelArena-α model, like TabPFN-Rel, in a few lines of code, all bundled as a simple PyPI package.
• Read the full model report: arxiv.org/abs/2608.16319
• GitHub repository (give us a ⭐️): github.com/PriorLabs/relaren…
• Announcement: priorlabs.ai/blog-posts/intr…
• Docs: docs.priorlabs.ai/capabiliti…
Thanks to the contributions from: @adrihayler, @KNfloege, @AlanArazi1536, @_rishabhranjan_, @jure, @LennartPurucker, @FrankRHutter & @noahholl
How would you let the LLM dream? [...] In a sense, to replay things, organize them, think about them so that tomorrow when we wake up, our thoughts are more organized than just input-output-input-output.
–Jure Leskovec, Stanford Professor, @jure