ReWorld: Learning Better Representations for World Action Models

Researchers have introduced ReWorld, a representation-learning framework that directly optimizes the intermediate representations inside world action models for autonomous driving, reporting gains in video generation quality and closed-loop planning performance [1]. World action models, or WAMs, predict how a driving scene will evolve under a given action and are seen as a scalable approach to autonomous-vehicle planning [1]. Until now, work on WAMs has concentrated on model architecture, while the question of how a model can efficiently learn internal representations that aid planning has received little attention [1]. Standard training supervises only the final outputs of the generation and planning modules; the intermediate representations that encode world knowledge are shaped only indirectly, as byproducts of fitting those outputs [1]. ReWorld treats those intermediate representations as direct optimization targets, shaping them along three complementary dimensions [1]. On the Video DiT responsible for generation, the method imposes future-predictive supervision on intermediate representations. On the Action DiT responsible for planning, it first aligns its representations cross-modally with the video world representation, then sharpens them around safety-critical boundaries using hard-negative supervision [1]. The authors also present a systematic analysis of existing representation-learning techniques applied to video-generation world models and explain why those methods deliver limited gains on this task [1]. The framework was evaluated on the nuScenes and NAVSIM datasets [1]. Fine-tuned video generation improved by 23.9 percent in Fréchet Video Distance, with the score moving from 81.3 to 61.9 [1]. Closed-loop PDMS rose from 89.1 to 90.4 without any post-training steps such as reinforcement learning or additional post-processing [1]. Convergence when training from scratch accelerated by roughly 2x [1]. The paper was posted on arXiv, an open-access repository for preprints that hosts work across computer science, physics, and other quantitative fields [6]. The broader context for this work sits within machine learning, a field in which statistical algorithms learn from data and generalize to unseen examples without being explicitly programmed [5]. The approach echoes a wider trend in artificial intelligence toward shaping internal representations directly, a strategy that proved transformative in natural language processing with models such as BERT, which learned contextual token representations through self-supervised objectives and became a ubiquitous baseline [4].