Obituary for the greatest monument to logical AGI.

New Preprint! Did you know that steering vectors from one LM can be transferred and re-used in another LM? We argue this is because token embeddings across LMs share many “global” and “local” geometric similarities! 1/N

#ICML #cognition #GrowAI We spent 2 years carefully curated every single experiment (i.e. object permanence, A-not-B task, visual cliff task) in this dataset (total: 1503 classic experiments spanning 12 core cognitive concepts). We spent another year to get 230 MLLMs evaluated

【描けるのに、”見えて”いない？画像生成AIのパラドックスに迫る】 生成AIを”見る”機械に変える「Diffusion Classifier」。しかし、精巧な絵を描けても、その内容を本当に「理解」しているのでしょうか？ ドイツのチュービンゲン大学などの研究は、最新のStable Diffusion

On the rankability of visual embeddings Ankit Sonthalia Arnas Uselis Seong Joon Oh tl;dr: one can discover "property ordering axis", such as age, etc in visual descriptors, often by having a couple of extreme examples. arxiv.org/abs/2507.03683

P.S. One of my fave Hopfield quotes. "How mind emerges from brain is to me the deepest question posed by our humanity." youtu.be/3OnjX-FtQOI?si…

🔍 How can we build AI agents that reason about the physical world the way humans do (or better) ? Excited to share Causal-PIK: Causality-based Physical Reasoning with a Physics-Informed Kernel, which will be presented next Thursday July 17th at ICML in Vancouver! 👇(1/6)

While making some figures for SGI* this year, I made some "behind the scenes" footage of how they get made: youtube.com/playlist?list=… Basically a video extension of cs.cmu.edu/~kmcrane/faq.h… ("figures?") *SGI is a great program run by Justin Solomon & deserves more funding

🧵0/7 🚨 Spotlight International Conference on Minority Languages 🚨 Chi-Ning and Hang have been thinking deeply about how feature learning reshapes neural manifolds, and what that tells us about generalization and inductive bias in brains and machines. They put together the thread below, which I’m sharing on

Test-time scaling nailed code & math—next stop: the real 3D world. 🌍 MindJourney pairs any VLM with a video-diffusion World Model, letting it explore an imagined scene before answering. One frame becomes a tour—and the tour leads to new SOTA in spatial reasoning. 🚀 🧵1/

Very excited to release a new blog post that formalizes what it means for data to be compositional, and shows how compositionality can exist at multiple scales. Early days, but I think there may be significant implications for AI. Check it out! ericelmoznino.github.io/blog/2025/08/1…

Does this mechanism apply to text-to-image models like SDXL? We find local/compositional structure in the learned feature-space — which helps predict the success of various compositions. But more work is needed on compositional generalization in large scale models!

alrighty, publicly sharing my slide deck for multimodal AI, covering ⤵️ > trends & uses > cool open-source models > tools to customize/deploy multimodal models > further resources all models in this presentation are on Hugging Face, easy load with 2 LoC!

Introducing Qwen3-VL Cookbooks! 🧑‍🍳 A curated collection of notebooks showcasing the power of Qwen3-VL—via both local deployment and API—across diverse multimodal use cases: ✅ Thinking with Images ✅ Computer-Use Agent ✅ Multimodal Coding ✅ Omni Recognition ✅ Advanced

🪩 Evaluate your LLMs on benchmarks like MMLU at 1% cost. In our new paper, we show that outputs on a small subset of test samples that maximise diversity in model responses are predictive of the full dataset performance. Project page: arubique.github.io/disco-site/ More below 🧵👇

Our part-aware 3D generation work, OmniPart, is accepted by Siggraph Asia 2025. Code and model released! Paper: arxiv.org/abs/2507.06165 Project page: omnipart.github.io Code: github.com/HKU-MMLab/Omni… Demo: huggingface.co/spaces/omnipar…

Children learn to manipulate the world by playing with toys — can robots do the same? 🧸🤖 We show that robots trained on 250 "toys" made of 4 shape primitives (🔵,🔶,🧱,💍) can generalize grasping to real objects. Jitendra MALIK trevordarrell Shankar Sastry Berkeley AI Research😊

Diffusion models learn probability densities by estimating the score with a neural network trained to denoise. What kind of representation arises within these networks, and how does this relate to the learned density? Eero Simoncelli Stephane Mallat and I explored this question.

Signature trait of LLM writing is that it’s low information, basically the opposite of this. You ask the model to write something and if you gloss over it you’re like huh okay this sounds decent but if you actually read it you realize half of the words aren’t saying anything.

CVPR keeps on scaring me with “ [Action needed CVPR 2026] Submission at risk of being desk-rejected” emails. Here is the latest one: Incomplete 3D Avatar Upload -> “Avatar Tasks” tab: •~Ulugbek_Kamilov