Do Independently Controlled Views Describe One Shared World?
MultiWorldBench evaluates independently controlled Minecraft views through navigation, shared-world interactions, and revisit consistency. This repository includes the V4 navigation protocol and the latest shared-world and revisit scoring code.
The accompanying manuscript is MultiWorldBench: Do Independently Controlled Views Describe One Shared World? (Zhangbo Xu, Ruoxi Zhang, Rui Hu, and Yisong Wang; October 2026).
The current release lives in benchmark_v4. Older controller code remains available for compatibility; its previous README is archived in docs/legacy-v3.md.
| Category | Case suite | Cases | Metrics |
|---|---|---|---|
| Navigation | Controlled motion | 180 | Control isolation; motion fairness; single-agent control consistency |
| Navigation | Cross-view translation | 20 | Cross-view motion consistency |
| Shared world | State synchronization | 120 | Shared-state synchronization |
| Shared world | State persistence | 50 | Shared-state persistence |
| Shared world | Building reveal | 75 | Cross-view structural coherence; spatial reasoning |
| Shared world | Concurrent interactions | 30 | Race-condition consistency |
| Revisit | Paired revisit / handoff | 20 | Style consistency; identity consistency |
There are 495 unique configured cases across seven suites. Configurations, metric implementations, portable scoring entry points, prompts, and regression tests are included. Single-agent consistency is also the eligibility diagnostic for navigation isolation.
This release evaluates supplied recordings and model outputs. Recorded videos, generated videos, model weights, training code, and the server-specific capture/inference orchestration are not bundled.
git clone https://github.com/Weather-etc/MultiWorldBench.git
cd MultiWorldBench/benchmark_v4
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python score_navigation.py --help
python score_shared_world.py --help
python score_revisit.py --helpSee the evaluation guide, input contracts, and metric definitions before preparing evidence.
Navigation uses map-anchored MCV camera poses with point replenishment, robust tracking, and optional backward recovery. V4 cross-view translation uses direct VLM judgments; its final case score requires both observer and actor judgments to be correct.
The deployed building geometry helper uses SAM2 plus image registration, including a camera-angle proxy. This helper does not implement the manuscript's SAM3/PnP geometry description; that difference is documented explicitly in the evaluation guide.
The release uses external MCV, SAM2, Qwen2.5-VL, and Cosmos-Reason1 components. The repository organization takes inspiration from WBench.
The exported evaluation code carries the source Apache-2.0 license in benchmark_v4/LICENSE. External dependencies and checkpoints retain their own licenses.