A new benchmark says today's AI agents still can't reliably turn a text prompt into a working 3D world.
Researchers built VibeWorlding, a framework for testing 'vibe worlding' agents: multimodal AI systems meant to infer what a user wants, plan a scene layout, pull in 3D assets, edit them, and render the result across multiple back-and-forth turns. The companion VWE-Bench dataset pairs 2,616 3D assets and 323 human-annotated seed worlds with 6,828 text queries generated by working backward from those scenes. On this benchmark, frontier models including GPT-5.5 and Qwen3.8-Max succeeded less than 60% of the time. The researchers pin most of the failure not on misreading the request, but on the models' inability to precisely edit a 3D scene once they've started building it.
That gap matters because it undercuts the polish of typical text-to-3D demos, which tend to showcase cherry-picked outputs rather than success rates across varied prompts. It also matters because the team's own reinforcement-learning-trained model, VibeWorlder-30B-A3B, ended up beating the closed frontier models on their own benchmark, a reminder that targeted post-training can still out-muscle raw scale on a narrow task.
Sub-60% is a long way from replacing a level designer. But it's a useful admission that generating a plausible-looking render is easy; editing it into what the user actually asked for is where these agents keep falling down.