VWE-Bench: Scoring Vibe Worlding Agents

VWE-Bench is the Vibe Worlding Evaluation Benchmark. It is a dataset and a scoring design, not a generator. The numbers to remember: 2,616 assets, 323 seed worlds, 6,828 queries.

Queries are reverse-synthesized from finished worlds, then split into verified items with ground truth and unverified items judged by rubrics. That mix is what lets the gym act as both an exam and a reward server for RL.

If you publish a new worlding agent, this is the first public yardstick aimed at end-to-end 3D construction rather than video FVD. If you just want a playable map, skip the bench and start with a blueprint.