🐙We are thrilled to share #Squidiff, a conditional diffusion model, which generates new transcriptomes that represent distinct cellular states, and its application to cell differentiation and drug perturbation tinyurl.com/mrxvyvn3 1/🐙

🚨 We’re honored to host Prof. Jürgen Schmidhuber (The Swiss AI Lab & KAUST) at the World Modeling Workshop 2026! ✨ A pioneer of modern AI, Prof. Jürgen Schmidhuber has made influential contributions that shaped the field — we’re thrilled to welcome him. 🌐 world-model-mila.github.io

I spent the past month reimplementing DeepMind’s Genie 3 world model from scratch Ended up making TinyWorlds, a 3M parameter world model capable of generating playable game environments demo below + everything I learned in thread (full repo at the end)👇🏼

Join us tomorrow @CoRL to discuss different views on simulations and world models in robotics with our amazing speakers! Final schedule & accepted papers links: simulatingrobotworlds.github.io/schedule.html 🤖🌏Room E8, starts at 9:30! 🌍🤖

🚀 ICLR rush is over—share your ideas at World Modeling Workshop 2026! 🧩 Posters, early results, negative findings, big ideas & published work welcome (non-archival). 📅 Abs: Oct 10 | 📅 Full: Oct 17 (AoE) 👉 bit.ly/wmw26-openrevi… #WMW2026 #WorldModeling

Great analysis

Exploration is fundamental to RL. Yet policy gradient methods often collapse: during training they fail to explore broadly, and converge into narrow, easily exploitable behaviors. The result is poor generalization, limited gains from test-time scaling, and brittleness on tasks

Introducing Scalable Option Learning (SOL☀️), a blazingly fast hierarchical RL algorithm that makes progress on long-horizon tasks and demonstrates positive scaling trends on the largely unsolved NetHack benchmark, when trained for 30 billion samples. Details, paper and code in >

Dylan Foster (Dylan Foster 🐢), Adam Block and I are giving a tutorial this year's NeurIPS '25! We'll be introducing the recent progress that has been made in understanding imitation learning, both in pretraining LLMs and in robotics. Check out Dylan's post for more!

🎮 How can agents learn to generalize from limited offline data? We introduce iMac (Imagined Autocurricula) - training agents entirely in world models with emergent curricula!

You have an ongoing project on world models? Want to meet fellow researchers and discuss your work? We extended our deadline by +2 weeks to the 1st November! (Abstract due on 25th Oct) Submit any work-in-progress negative results, or already published work ➡️

Diffusion models operate step-by-step, hence they are serial models right? However, that doesn't sit well because we have been seeing that diffusion models don't scale well with "steps". In this thread: Diffusion models are not truly serial models.

🔊Simplicial Embeddings (SEMs) Improve Sample Efficiency in Actor-Critic Agents🔊 In our recent preprint we demonstrate that the use of well-structured representations (SEMs) can dramatically improve sample efficiency in RL agents. 1/X

Congratulations to Aishwarya Agrawal (Aishwarya Agrawal), Core Academic Member at Mila, who has received today the Everingham Prize 2025, one of the most prestigious awards in the computer vision community, in recognition of her foundational contributions to the field.

🚨The Formalism-Implementation Gap in RL research🚨 Lots of progress in RL research over last 10 years, but too much performance-driven => overfitting to benchmarks (like the ALE). 1⃣ Let's advance science of RL 2⃣ Let's be explicit about how benchmarks map to formalism 1/X

🚨 Interested in generative world models? We’re thrilled to host Stephen Spenser (Google DeepMind) at the World Modeling Workshop 2026, where he’ll talk about the Genie series of models! 🌐 world-model-mila.github.io

Our paper "Vision Transformers Don't Need Trained Registers" will appear as a Spotlight at NeurIPS 2025! We uncover the mechanism behind high-norm tokens and attention sinks in ViTs, propose a training-free fix, and recently added an analytical model -- more on that below. ⬇️

#EMNLP2025 has kicked off, and it's special or us this time! 📌 𝐂𝐎𝐋𝐌𝐀𝐓𝐄: tackles the long-standing enterprise nightmare of document search & retrieval 👇 📌 𝐖𝐞𝐛𝐌𝐌𝐔: pushes web UI reasoning, visual code generation & editing across languages. arxiv.org/abs/2508.1676

Couldn’t be prouder of what this team pulled off — small model, big leap for grounded agents. Everything’s open, can’t wait to see what others build with it! 🔥

I do not think you can pursue meaningful research without (1) some grandiose delusion about your abilities (2) a sense of esthetics and harmony to judge ideas still free of experimental confirmation (3) an unreasonable taste for the required tangible work (e.g. programming)