Tal Daniel

🚀 #ICLR2026 Oral 💥 How can we design world models that capture object interactions directly from pixels? Introducing Latent Particle World Models-the first end-to-end self-supervised, object-centric world model, trained from videos, supporting action/img/lang conditioning. 1/n

@DanHrmti @carl_qi98 @LambdaAPI @AvivTamar1 @davheld @pathak2206 🌐 Learn more about Latent Particle World Models (LPWM): Website: taldatech.github.io/lpwm-web/ Paper: arxiv.org/abs/2603.04553 Code (+ tutorial): github.com/taldatech/lpwm 16/n

Huge thanks to @DanHrmti, @carl_qi98, Amir Zadeh and Chuan Li (@LambdaAPI), @AvivTamar1, @davheld, and @pathak2206 — incredible collaborators who made LPWM possible. 15/n

During training, the posterior latent actions condition the dynamics module that predicts the next-frame prior. A KL regularization term aligns this prediction with the latent policy’s output, forming a VAE-style objective over particle transitions. 7/n

The inverse dynamics observes particles at t and t+1, inferring the latent actions that caused the change. The latent policy sees only particles at t and outputs a distribution over possible latent actions from the current state. 6/n

🚀 #ICLR2026 Oral 💥 How can we design world models that capture object interactions directly from pixels? Introducing Latent Particle World Models-the first end-to-end self-supervised, object-centric world model, trained from videos, supporting action/img/lang conditioning. 1/n

To address this, we introduce a context module that predicts latent actions per particle, enabling fine-grained, multi-entity dynamics It has two heads: (1) an inverse dynamics (posterior) and (2) a latent policy (prior). 5/n

Building a world model means capturing stochastic particle dynamics. Existing “latent action” models help, but (1) need strong regularization (e.g., VQ) and (2) rely on a single global latent—missing interactions among multiple entities. 4/n

DLP decomposes scenes to particles with several attributes (keypoints, bounding-boxes, masks), fully unsupervised. These act as visual “tokens,” making cross-modal long-horizon reasoning (vision ↔ language) far more natural than the standard pixel patches. 3/n

Self-supervised object-centric models decompose scenes into entities without any supervision—a key step toward visual understanding. In this work, we extend Deep Latent Particles (DLP) to world-modeling on real-world datasets. 2/n

Code, datasets and checkpoints are now available! github.com/DanHrmti/HECRL #WorldModels #Diffusers #OfflineGCRL #HierarchicalRL #EntityCentricRL

Check out our latest work on HRL and (self-supervised) object-centric representations! Led by the talented @DanHrmti

Kudos to the amazing hard-working organization team @lucasmaes_ , @hafezghm , Yipeng, and for @LambdaAPI for their generous support of the workshop.

We have an amazing line-up of rockstar speakers, and we will also debut our “Latent Particle World Models” work (Thursday Feb 5, 11am EST). Schedule and livestream links: world-model-mila.github.io

Interested in world models? The first “World Modeling” workshop is happening this week (Feb 4-6, Wed-Fri) at @Mila_Quebec , and it will be streamed on YouTube so you can follow along live!