Reproduce “PlayWorld: Benchmarking World Models with Agent Players over Long-Horizon Objectives” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Kaixin Ding, Xi Chen, Minghong Cai, Zhiyuan Xu, Yiyang Wang, Yuxiang Lu, Junyi Li, Shuyang Chen and 4 more

PlayWorld introduces a benchmark for evaluating interactive video world models by employing multi-modal Agent Players to pursue long-horizon objectives and adaptively adjust action sequences. The evaluation, performed on nine models, indicates that current systems face significant challenges in maintaining sustained world evolution and global spatial consistency across complex interactions.

Key insights

  • An adaptive control strategy, combining a shared action prior with online, observation-conditioned adjustments by an agent, reliably guides world models toward long-horizon objectives.
  • Current world models struggle with sustained world evolution, consistently showing low performance in maintaining semantic state changes over extended periods or across occlusions.
  • Long-horizon revisitation tasks expose global spatial inconsistencies, as models frequently fail to preserve unique spatial locations of landmarks, regenerating them inconsistently.
  • Interaction fidelity remains limited beyond basic collision detection, with many models failing to produce physically plausible visual or physical responses to complex actions like water entry.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.