Are sota image generative models effective world models? Consistency-diversity-realism Pareto fronts show they're not (yet): - No model dominates others as a world model - Improvements in quality and consistency have come at the expense of diversity 🔗 arxiv.org/abs/2406.10429

Why do LLMs fail simple completion tasks, but not on a harder task? Learn about tokenization bias in LLMs and how to fix it by compute exact string probability! arxiv.org/pdf/2406.16829 work w/Marton Havasi Matthew Muckley Dr. Karen Ullrich accepted in Theoretical Foundations of Foundation Models, ICML2024 (1/4)

Appearing at ICML this week! icml.cc/virtual/2024/p…

Even with preference alignment, LLMs can be enticed into harmful behavior via adversarial prompts 😈. 🚨 Breaking: our theoretical findings confirm: LLM alignment is fundamentally limited! More details, on framework, statistical bounds and phenomenal defense results 👇🏻

Tokenization is a limitation of modern LLMs. How to make it better? Find a way to convert your LLM's probabilities to byte-level probabilities! Details in the thread below! ⬇️ Work done by our great intern Buu Phan

If you prompt an LLM and stop in the middle of a token, what happens? ❌ The generated response doesn't correctly complete the token 📚 Our new paper provably fixes this for better code generation and ensembling! arxiv.org/abs/2410.09303 More details in Buu Phan's thread below

Here we go again...

Meta just released V-JEPA 2: new open-source image/video world models ⏯️🤗 > based on ViT, different sizes (L/G/H) and resolution (286/384) > 0-day support in 🤗 transformers > comes with a physical reasoning (from video) benchmark: MVPBench, IntPhys 2, and CausalVQA

Say hello to DINOv3 🦖🦖🦖 A major release that raises the bar of self-supervised vision foundation models. With stunning high-resolution dense features, it’s a game-changer for vision tasks! We scaled model size and training data, but here's what makes it special 👇

(🧵) Today, we release Meta Code World Model (CWM), a 32-billion-parameter dense LLM that enables novel research on improving code generation through agentic reasoning and planning with world models. ai.meta.com/research/publi…

1/ My model can’t generate “full” refrigerators when I ask for it?? — but it does by default anyway when I don’t specify!? At #ICCV2025, we’re presenting DIMCIM — a new framework to quantify diversity and capacity in T2I models, that enables pinpointing such failure modes