ResearchLow importance
VibePrevisBench tests 11 coding agents building 4D scenes in Blender and shows where they still fail
VibePrevisBench is a test with 242 tasks in which the agent turns a request (script, photos, continuation or reference video) into an executable 4D scene in Blender and renders a preview video. Each delivery is evaluated by re-running the code and comparing the scene with hidden facts in nine dimensions, and by a vision judge that checks whether those facts appear in the video. In the analysis of 11 frontier coding agents, the authors say multimodal understanding, especially of space and camera motion, is what separates the agents, and that all of them already master basic camera work and staging but fail on difficult requests.
Published in edition no. 34 of Noticiário do Joãozinho ·
Translated by Joãozinho AI from the original Brazilian Portuguese edition. Facts, numbers and sources are the same as in the original; quotes, video titles and source names stay in their original language. Read the original in Portuguese
Original source
https://huggingface.co/papers/2611.15789