Title: WorldLens: Comprehensive Real-World Assessment of Driving World Models

Abstract:

Generative world models are fundamentally transforming embodied AI by allowing agents to construct convincing 4D driving simulations. However, while these environments often appear visually plausible, they frequently fall short in terms of physical accuracy and behavioral logic. To date, the field has lacked a standardized framework for determining whether generated worlds maintain geometric integrity, adhere to physical laws, or facilitate dependable control.

In response, we present WorldLens, a holistic benchmark designed to evaluate how effectively models construct, comprehend, and operate within their synthetic environments. This assessment covers five critical dimensions: Generation, Reconstruction, Action-Following, Downstream Task performance, and Human Preference. Collectively, these categories address visual realism, geometric consistency, physical plausibility, and functional reliability.

Our analysis reveals that no current world model demonstrates universal superiority. For instance, models that excel in rendering high-quality textures often compromise on physical accuracy, whereas those that maintain geometric stability tend to lack behavioral fidelity.

To bridge the gap between automated metrics and human perception, we introduce WorldLens-26K, a large-scale dataset comprising human-annotated videos accompanied by numerical ratings and textual justifications. Additionally, we developed WorldLens-Agent, an evaluation model distilled from these annotations to facilitate scalable and explainable scoring. Together, this benchmark, dataset, and agent create a cohesive ecosystem for measuring world fidelity, establishing a new standard for judging future models not merely on their visual appearance, but on the authenticity of their behavior.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC