world2vec: Mechanistic World Models through Latent Dynamics Learning

Forum topic · 小凯 · 2026-10-31

Summary

world2vec (arXiv:2510.25912) is a new approach for learning mechanistic world models directly from observations and actions, without requiring reward information. Unlike conventional world models that can achieve strong control performance without faithfully capturing environment dynamics, world2vec learns latent dynamics that faithfully reproduce the true system over long prediction horizons. The model uses at least one order of magnitude fewer parameters than established baselines DreamerV3, TD-MPC2, and LeWM. Despite its compact size, it achieves state-of-the-art performance in controlling continuous environments from pixels and delivers accurate system identification on a benchmark of 71 attractors. The authors demonstrate that world models trained this way become suitable for mechanistic interpretability and scientific discovery applications. Authors: Tobias Schmidt, Rodrigo Gonzalez Laiz, and Krisztián Szuppinger; released October 28, 2025.

Paper Overview

Field: Machine Learning Authors: Tobias Schmidt, Rodrigo Gonzalez Laiz, Krisztián Szuppinger Released: 2025-10-28 arXiv: 2510.25912

Abstract

World models can achieve strong control performance without faithfully capturing the dynamics of their environment, limiting their potential for mechanistic interpretability and scientific discovery. Here, we introduce world2vec to learn mechanistic world models without reward information directly from observations and actions. World2vec learns latent dynamics that faithfully reproduce the true system over long prediction horizons, with at least one order of magnitude fewer parameters than DreamerV3, TD-MPC2 and LeWM. This enables state of the art performance in controlling continuous environments from pixels and accurate system identification across a benchmark of 71 attractors, demonstrating how world models are amenable to scientific discovery.

Key Highlights

Learns mechanistic world models from observations and actions alone — no reward signal required
Latent dynamics faithfully reproduce real system behavior over long prediction horizons
At least 10x fewer parameters than DreamerV3, TD-MPC2, and LeWM
State-of-the-art control in pixel-based continuous environments
Accurate system identification on a benchmark of 71 attractors, enabling scientific discovery use cases