Latent Dynamics · 2026-10-30
world2vec: Mechanistic world models through latent dynamics learning
Tobias Schmidt, Rodrigo Gonzalez Laiz, Kriszti\'an Szuppinger, Steffen Schneider
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
World2vec learns mechanistic world models from observations and actions without using reward information. It learns latent dynamics that reproduce the true system over long prediction horizons, and the authors report “at least one order of magnitude fewer parameters” than DreamerV3, TD-MPC2, and LeWM. The method achieves state of the art control on continuous environments from pixels and accurate system identification across a benchmark of 71 attractors.
Why it matters
The authors target mechanistic interpretability by learning faithful latent dynamics over long horizons, rather than only optimizing control rewards. They frame a dynamics matching objective that aims for latent rollouts to match encoded future observations, and they report both strong control and system identification results, supporting the claim that world models can support scientific discovery.
Method
- Learns an encoder h mapping a short observation history to latents z and an action-conditioned neural ODE/vector-field f̂; training uses latent dynamics matching over multi-step horizons with explicit integration (Euler).
- Uses mechanistic faithfulness conditions formulated in terms of invertible coordinate changes and matching latent trajectories produced by integrating the learned latent velocity field.
- For control, performs MPC in latent space from observation-only goal sequences, using CEM planning and distances to a goal latent; no learned reward model or state decoder is used.
Abstract (from arXiv)
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.
Related papers
- Structured Latent Dynamics for Long-Horizon Embodied World Models
- $\Sigma$-LeWM: Learning World Models via S-Information for Robot Control
- Understanding the limits of latent world models
- MemWM: Associative Memory-Enhanced Latent World Models for Long-Horizon Planning
- K-PWM: Control-Oriented Structured Bilinear World Models under Partial Observation
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-29 · Indexed by Pub-AI: 2026-10-30 UTC
world2vec: Mechanistic world models through latent dynamics learning
Preprint posted: 2026-10-29 · Original record ↗
- arxiv
- 2610.40637
Explicit source versions (1)
- v1 · 2026-10-29
- Preprint available
Source date: 2026-10-29 · First detected: 2026-10-30 UTC · Historical baseline
Source evidence (2)
- Official evidence ↗ · RSS guid=oai:arXiv.org:2610.40637v1; arxiv:announce_type=new; item pubDate (official metadata)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
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.
