VibeWorlding benchmark: open 30B model tops GPT-5.5 on 3D worlds

Below 60%. That is where GPT-5.5 and Qwen3.8-Max land on VWE-BENCH, a new evaluation aimed at whether multimodal agents can build interactive 3D open worlds end to end from a natural-language prompt.

The benchmark comes from the paper VibeWorlding: Can Multimodal Agents Construct 3D Open Worlds End-to-End? by Yansong Ning and colleagues. They assembled "a benchmark of 2,616 high-quality 3D assets, 323 human-annotated seed 3D worlds, and 6,828 reverse-synthesized multimodal user queries," then measured Pass@1 on the full pipeline.

Their read on the frontier models is blunt: "current frontier MLLMs are far from solving the vibe worlding agent task, with even GPT-5.5 and Qwen3.8-Max reaching below 60% success rate." The paper points at "precise 3D world editing" as the primary bottleneck.

The counterpunch is an open-weight 30B model. VibeWorlder-30B-A3B, post-trained inside the authors' VibeWorlding-Gym with sandbox tools and verifiers, "attains the best overall Pass@1 among all evaluated models." The gain is credited to reinforcement learning against verifiable rewards rather than raw parameter scale.

No per-task breakdowns or annotator agreement figures appear in the abstract, and every number here is the authors' own scoring on their own benchmark.