Reproduce “Variational Causal Dynamics: Discovering Modular World Models from Interventions” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Anson Lei, Bernhard Schölkopf, Ingmar Posner

Variational Causal Dynamics (VCD) jointly learns latent representations, sparse temporal causal graphs, and environment-specific mechanism changes from state or image observations. In simulated multi-body systems, it matched environment-specific baselines on mixed-state prediction, outperformed shared and separately trained alternatives on image rollouts, and adapted more accurately with limited data by updating only inferred intervened mechanisms.

Key insights

  • The causal objective encourages latent dimensions to align more closely with individual physical variables: in mixed-state experiments, perturbing a VCD latent generally affected one ground-truth variable rather than several.
  • Temporal causal structure avoids the need for within-time-step acyclicity, allowing each previous-state variable to influence any next-state mechanism while preserving modularity.
  • Sparse mechanism reuse provides a middle ground between one model shared across all environments and fully separate models, improving data efficiency under localized distribution shifts.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.