AION
Research paperRobotics & Embodied AI · Multimodal Models · Safety & Alignment1 source · Oct 8, 2026

MultiWorldBench: Do Independently Controlled Views Describe One Shared World?

We introduce MultiWorldBench, a diagnostic Minecraft benchmark containing 495 case configurations across seven task suites and ten capabilities, including independent control, cross-view motion, shared-state synchronization, persistence, structural reasoning, concurrent interaction, and delayed revisit.

Key points

  • Multiplayer world models must ensure that independently controlled views remain consistent with one shared and persistent world.
  • We evaluate Solaris, Gamma-World, and MineWorld, using Engine GT as a reference.
  • Gamma-World performs better on several control, shared-state, and revisit capabilities, whereas Solaris leads in cross-view motion and race-condition consistency.
  • These results show that plausible individual views do not yet constitute a coherent multiplayer world.

Sources (1)

  • [1]MultiWorldBench: Do Independently Controlled Views Describe One Shared World?
    arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 8, 11:24 AM
    We introduce MultiWorldBench, a diagnostic Minecraft benchmark containing 495 case configurations across seven task suites and ten capabilities, including independent control, cross-view motion, shared-state synchronization, persistence, structural reasoning, concurrent interaction, and delayed revisit.
    Multiplayer world models must ensure that independently controlled views remain consistent with one shared and persistent world.

Extractive summary: sentences quoted from the sources.