World models learn true dynamics with fewer parameters

A new method called world2vec reproduces environmental dynamics over long horizons using a tenth the parameters of prior models, enabling accurate control from pixels and scientific discovery.

Research Lab

Tobias Schmidt · Rodrigo Gonzalez Laiz · Krisztián Szuppinger · Steffen Schneider

Helmholtz Munich · Munich Center for Machine Learning (MCML)

Research Digest·30 October 2026·2 min read

Illustration: World models learn true dynamics with fewer parameters

In one line

World2vec learns faithful, compact world models from observations and actions without rewards, matching reward-trained control and identifying dynamical systems.

Before you read this

Three ideas this paper builds on

world model for agentsA predictive environment model that agents use to simulate outcomes, enabling planning, reasoning, and long-horizon task execution.93 papers →mechanistic interpretabilityThe study of how neural networks compute through internal circuits, features, and causal interventions, rather than just input-output behavior.98 papers →model predictive controlA control method that uses a model to predict future outcomes and selects actions to optimize performance over a horizon.4 papers →

Read it as

To control a robot, software needs a model of how the world behaves. Most learned models are built around rewards, so they act well without truly understanding the dynamics. World2vec builds a world model from observations and actions alone, with no reward: a compact latent representation and a dynamics model that predicts how hidden state evolves.

World2vec is trained by predicting future observations, not by being told the right answer. Across a benchmark of 71 attractors, systems that settle into repeating patterns, it recovered the true underlying dynamics. It also matched or beat reward-trained models on control from pixels while using far fewer parameters.

This matters for science: a faithful model is read out as dynamics, not just a controller.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

world model for agents mechanistic interpretability latent dynamics learning model predictive control joint-embedding predictive architecture dynamics contrastive learning