🤔 How do we train AI models that surpass their teachers? 🚨 In #COLM2025: ✨Delta learning ✨makes LLM post-training cheap and easy – with only weak data, we beat open 8B SOTA 🤯 The secret? Learn from the *differences* in weak data pairs! 📜 arxiv.org/abs/2507.06187 🧵 below

What if the secret to improving an LLM isn’t with better data or a better teacher? The Delta Learning Hypothesis: learn from the **delta** in performance between two weaker LLMs to instruction-tune a stronger one.

Large foundation models trained on massive datasets have revolutionized #AI. Supported by a Google Research Ph.D. Fellowship, University of Washington #UWAllen’s Cheng-Yu Hsieh aims to make the process more efficient and affordable to democratize AI development. #UWdiscovers news.cs.washington.edu/2025/07/09/all…

“Technical computer science savvy and deep philosophical commitments”: University of Washington #UWAllen alum Andre Ye was named the UW College of Arts & Sciences Dean’s Medalist in Social Sciences for his campus leadership and research contributions spanning #AI and philosophy. #UWdiscovers artsci.washington.edu/news/2025-06/2…

🥳 Gap year update: I'll be joining Ai2/University of Washington for 1 year (Sep2025-Jul2026 -> JHU Computer Science) & looking forward to working with amazing folks there, incl. Ranjay Krishna, Hanna Hajishirzi, Ali Farhadi. 🚨 I’ll also be recruiting PhD students for my group at JHU Computer Science for Fall

If you are a near-graduation PhD student in computer vision, consider applying to the ICCV 2025 Doctoral Consortium (DC). It is a chance to be mentored by an experienced researcher in the vision community to help you transition to your post-PhD career in academia or industry.

🧵Excited to share our new paper: MedBLINK 🩻 Would you trust ChatGPT with your X-ray if it couldn't tell if the image is upside down? We introduce MedBLINK, a benchmark that evaluates MLMs on basic perception tasks that are trivial for clinicians but often fail for AI.

Today's foundation models can't tell if a CT scan is upside down. Or if it is looking at a pediatric or adult x-ray scan... so before you start asking it to diagnose your condition, think twice!

🎨🤖 Call for Submissions! Join us at Humanoids 2025 in Seoul for Embodied Co‑Creation: Robotic Tools That Make and Perform Arts — a unique platform where robotics meets creativity! 📅 Oct 2 | COEX Seoul 🌐 embodied-co-create.github.io

🚨 Tested GPT-5 on our MedBLINK benchmark: 76.3% average accuracy, +12.3% over the previous best (Claude 3.5 Sonnet). Strong improvement, yet still 20% behind humans (96.4%) on simple, basic medical perception tasks; the gap remains wide.

MolmoAct Action Reasoning Models that can Reason in Space

MolmoAct: Action Reasoning Models that can Reason in Space depth → trajectory → actions - Backbone: Molmo VLM (OpenCLIP/OLMo2-7B or SigLIP2/Qwen2.5-7B) + ordinal action tokens (256 bins, 5.4× less pretrain compute) - Data: 10.6k Franka trajectories (93 tasks) + OXE subset

MolmoAct: Action Reasoning Models that can Reason in Space "Reasoning is central to purposeful action, yet most robotic foundation models map perception and instructions directly to control, which limits adaptability, generalization, and semantic grounding. We introduce

Reasoning is central to purposeful action. Today we introduce MolmoAct — a fully open Action Reasoning Model (ARM) for robotics. Grounded in large-scale pre-training with action reasoning data, every predicted action is interpretable and user-steerable via visual trace. We are

Introducing MolmoAct Our Action Reasoning Model that can Reason in Space (1/)🧵⬇️

We are launching MolmoAct🤖✨ A fully open Action Reasoning Model (ARM) that can reason in space: it perceives → it plans → it acts. 🧵👇

✨Thrilled to see our perception tokens used in robotics: MolmoAct predicts depth tokens first, then plans trajectories and actions. Love this direction for grounded action reasoning. check out the perception tokens here: aurora-perception.github.io

This to me really feels like how robot foundation models "should" work. i like that it can autoregressively predict depth tokens, lift to 2.5d, and use this for reasoning - it feels like a true robotics analogue of modern reasoning LLMs. Really exciting work.

AI2's MolmoAct model ‘thinks in 3D’ to challenge Nvidia and Google in robotics AI venturebeat.com/ai/ai2s-molmoa…

Most AI models still think in words. People, without even noticing, think with our bodies, planning how to move, grasp, and use things around us. MolmoAct brings that to robotics: reasoning in space before acting. This is how we will get to the GPT-moment for robotics.