Driving · 2026-11-02
CausalWAM: Learning the Causes of Actions in World Action Models
Juntao Cheng, Jinzhou Tang, Fan Feng, Henrik I. Christensen, Yujia Zheng
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
CausalWAM addresses a mismatch in world action models: predicting the whole scene is not the same as learning which latent causes drive actions. The authors prove that, with a sparsity constraint, action-relevant latent variables are identified up to permutation and the deployed compact state is an invertible function of the true action-relevant block. They propose CWAM by adding a sparse routing path to the action head of Cosmos 3. Experiments validate theory and improve closed-loop success efficiency.
Why it matters
World action models can fit demonstrations while entangling action-relevant and nuisance latents, so action generation may depend on variables that only help scene prediction. The paper’s identifiability results separate “what causes actions” from “what prediction depends on” using only observational demonstrations. The proposed sparse-routing implementation keeps the world-modeling objective unchanged and targets the action side, aiming to improve action reliability in closed loop.
Method
- Prove that without a support-sparsity constraint, observational fitting only guarantees an inequality on action-side Jacobian support; sparsity turns it into equality and identifies the latent-to-action dependency pattern up to permutation.
- Under the same hypotheses, prove the compact deployed action-relevant state C_t is an invertible re-encoding of the true action-relevant latent block, even if action-relevant latents are statistically dependent on nuisance latents.
- Implement CWAM (Causal WAM) by adding a sparse routing path to a pretrained WAM’s action head (Cosmos 3), applying a token-level surrogate sparsity loss to the router readout while keeping the world-modeling objective unchanged.
Limitation
The span condition in Proposition 1 is explicitly assumed: the Jacobian rows of each action dimension must span the coordinate subspace indexed by its true parents, and smoothness/positive state density alone do not imply it.
Abstract (from arXiv)
World action models (WAMs) learn latent variables from observations and use these to both predict the evolution of the scene and to generate actions. The prediction has to be of the whole scene, whereas only a portion of the scene causes actions. WAMs fitted to demonstrations do not, in general, guarantee that they will generate actions solely from the latent causes of the demonstrated actions (the action-relevant latent variables), as the model could just as well learn to imitate in this way whilst entangling the action-relevant latent variables with the nuisance variables. We therefore separate out what causes action from what prediction depends on. We prove that under a sparsity constraint the latent variables used to generate actions are an invertible function of the action-relevant latent variables alone and the true structure is identified up to a permutation of the latent variables. This requires neither sufficient temporal dependence nor any assumptions on how the unobserved regimes evolve and holds even when the action-relevant latent variables are statistically dependent on the rest. We therefore propose CWAM (Causal WAM), which incorporates the constraint into a pretrained WAM by adding a sparse routing path to its action head, with the world modeling objective remaining unchanged. Experiments on both synthetic data and real-world robots validate the theory as well as showcase the improved efficiency in closed-loop success.
Related papers
- OpenWAM: An Open Framework for Composable World-Action Models
- {\Delta}WAM: Distilling Action Tangent Fields into World Action Models
- Long-WAM: Scaling the Context of World-Action Models
- CAG-WAM: Causal Action-Gain World Action Models for Robot Manipulation
- EVO-WAM: Evolving World Action Models through Video-Action Verification
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-30 · Indexed by Pub-AI: 2026-11-02 UTC
CausalWAM: Learning the Causes of Actions in World Action Models
Preprint posted: 2026-10-30 · Original record ↗
- arxiv
- 2610.42096
Explicit source versions (1)
- v1 · 2026-10-30
- Preprint available
Source date: 2026-10-30 · First detected: 2026-11-02 UTC · Historical baseline
Source evidence (3)
- Official evidence ↗ · RSS guid=oai:arXiv.org:2610.42096v1; arxiv:announce_type=new; item pubDate (official metadata)
- Official evidence ↗ · atom:published (official metadata)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
World action models (WAMs) learn latent variables from observations and use these to both predict the evolution of the scene and to generate actions. The prediction has to be of the whole scene, whereas only a portion of the scene causes actions. WAMs fitted to demonstrations do not, in general, guarantee that they will generate actions solely from the latent causes of the demonstrated actions (the action-relevant latent variables), as the model could just as well learn to imitate in this way whilst entangling the action-relevant latent variables with the nuisance variables. We therefore separate out what causes action from what prediction depends on. We prove that under a sparsity constraint the latent variables used to generate actions are an invertible function of the action-relevant latent variables alone and the true structure is identified up to a permutation of the latent variables. This requires neither sufficient temporal dependence nor any assumptions on how the unobserved regimes evolve and holds even when the action-relevant latent variables are statistically dependent on the rest. We therefore propose CWAM (Causal WAM), which incorporates the constraint into a pretrained WAM by adding a sparse routing path to its action head, with the world modeling objective remaining unchanged. Experiments on both synthetic data and real-world robots validate the theory as well as showcase the improved efficiency in closed-loop success.
