Skip to main content
Search

# Programmable World Model

1 related articles

Subscribe to updates
  1. arXiv

    PROWBench tests video models on 170 replayable programmatic worlds: proxy-conditioned models place entities better, but 30-second horizons and multi-view identity binding still break down

    The work introduces PROWBench, comprising 170 programmatically constructed episodes, 200 selected camera views and 600 synchronized proxy videos, using replayable world states and timestamped engine events as ground truth to evaluate whether video models faithfully render program-specified scene structure, rules and interactions, and proposes two VLM-based metrics, Logic-Render Alignment and Interaction Success Rate; results show proxy-conditioned models clearly outperform camera-conditioned world models on entity control, while long-horizon memory, multi-view identity and appearance binding remain open problems.

You’re all caught up