VibeWorlding Benchmark: Open 30B Model Beats GPT-5.5 on 3D World Building
Justgototheeffinmoon · reddit · 2026-08-18
A new paper, "VibeWorlding," introduces VWE-BENCH to evaluate if multimodal agents can build interactive 3D open worlds end-to-end from text prompts. The study finds that frontier models like GPT-5.5 and Qwen3.8-Max struggle, with success rates below 60% due to difficulties in precise 3D editing.
Surprisingly, an open 30B parameter model, VibeWorlder-30B-A3B, achieved the best overall Pass@1. The model was post-trained with sandbox tools and verifiers, suggesting that reinforcement learning with verifiable rewards trumps raw parameter scale for this task.
Related event: Tencent's VibeWorlding: Agents Build 3D Open Worlds End-to-End(2 posts)→
More from Multimodal
- How meloform works: WASM pitch detection plus differentiable circuits learning 1,022 sine coefficients — henloitsjoyce · 2026-10-16
- Documentary filmmaker urges peers to watch 'You Can See Everything' — kyliebytes · 2026-10-16
- Redditor's first Cumfy/Seedance video nails a cinematic one-shot generation — Wave_N_Wind · 2026-10-16
- Kling 4 returns after 8 months: 10 hyper-realistic AI ads for under $5 — CurieuxExplorer · 2026-10-16
- Shanghai AI Lab unveils Intern Lumina U2, a diffusion LLM unifying image, video and 3D tasks — Shanghai-AI-Laboratory · 2026-10-16
- NAVER AI Lab proposes Speculative Forcing to speed up autoregressive video generation — naver-ai · 2026-10-16