VibeWorlding-Gym and VWE-Bench Explained

VibeWorlding-Gym is the training environment. VWE-Bench is the exam. Both shipped on 16 August 2026 from HKUST USAIL with the VibeWorlder model family.

The gym gives the agent a sandbox: retrieve an asset, edit it, render views, all as MCP tools. The verifier mixes physical checks (did two meshes collide?) with intent checks (did you delete the statue closest to the altar, not just any statue?). That second test is spatial, not lexical. The paper uses identical statue assets on purpose.

What is in VWE-Bench

2,616 3D assets, 323 human-annotated seed worlds, and 6,828 reverse synthesized multimodal queries. Some queries have ground truth. Others use rubrics. The split lets researchers train with a reward service instead of only grading a static test set.

If you are shopping for a consumer world generator, this stack is research infrastructure. It tells you whether an agent can follow world-edit instructions. It does not replace a game runtime. Use it to read the field. Use the blueprint tool when you want a brief you can play.