Fast LeWorldModel

A new latent world model called Fast LeWorldModel accelerates visual planning by replacing the step-by-step simulation used in its predecessor with a parallel action-prefix prediction method, according to a paper submitted to arXiv on 24 Jun 2026 [1]. The predecessor, LeWorldModel, belongs to a family of Joint-Embedding Predictive Architectures that build internal representations of environments without reconstructing raw images [1][2]. World models of this kind simulate dynamics such as physics and object interactions, allowing artificial agents to plan actions without constant real-world trial and error [6]. LeWorldModel evaluates candidate action sequences through an autoregressive rollout, applying a local one-step latent transition model repeatedly [1][3]. This approach becomes computationally expensive and exposes predicted trajectories to accumulating latent errors as the planning horizon lengthens [1][4]. Fast LeWorldModel addresses those limitations by encoding action prefixes and predicting the future latent states reached after executing those prefixes, all in a single parallel forward pass [2][5]. The model uses an action-prefix encoder with a causal mask to convert a candidate action sequence into multiple prefix tokens, where each token summarizes a partial sequence together with the current latent context [3][5]. A parallel latent predictor then maps the current latent and all prefix tokens to their corresponding future latents simultaneously [3][5]. Dense prefix-level supervision ties each prefix token to the future latent reached after executing the corresponding partial action sequence, forcing the model to learn how states continuously evolve under different action prefixes rather than only fitting one-step transitions [2][3]. During planning, the predictor can use the last prefix token from the encoded action sequence to evaluate the corresponding future latent without explicitly rolling through each intermediate imagined state [1][4]. Across multiple tasks, Fast LeWorldModel improved average success over LeWorldModel while substantially reducing planning time [1][2]. It also achieved lower open-loop latent loss, with loss growth becoming significantly slower as the rollout horizon increased [1][4]. The paper was posted on arXiv, a repository that provides rapid dissemination of research findings at no cost to readers and submitters [10].