NeFut Logo NeFut
中 Admin Login

[CS.AI] OneWorld: Learning Consistent Physics Across Actions in World Models

Published at: 2026-09-29 22:00 Last updated: 2026-09-30 01:41
#algorithm #AI #Machine Learning

Action‑conditioned video world models aim to forecast scene evolution under different actions, a capability essential for reliable planning and decision‑making. Yet independent futures generated from the same initial scene can each look plausible while encoding conflicting physical properties such as friction or mass, leading to cross‑intervention inconsistency and reduced trustworthiness. To tackle this, we introduce OneWorld, a shared‑mechanism counterfactual generation framework that jointly models multiple action‑conditioned futures under a common latent physical mechanism. A physical‑mechanism interpreter first infers a distribution over latent mechanisms for each action‑outcome branch; these distributions are then aggregated into shared‑world evidence that measures whether the branches admit a unified physical explanation while preserving uncertainty in less informative branches. This evidence constrains flow training and guides sampling, encouraging consistency of the underlying physics while retaining distinct outcomes of different actions. We also propose a multi‑intervention evaluation protocol in controlled environments, following the ACWM‑Phys interaction settings, to test whether generated futures can be jointly explained by the same physical parameters alongside standard single‑rollout prediction metrics. Experiments show that OneWorld improves cross‑intervention physical consistency while keeping single‑rollout performance competitive.

Review: OneWorld’s shared latent mechanism yields consistent physics across actions, offering a promising direction for more interpretable and reliable world models.

Original Source: https://arxiv.org/abs/2609.30946

[h] Back to Home