This is a Plain English Papers summary of a research paper called VibePrevisBench Tests Whether Coding Agents Can Build Readable 4D Blender Scenes. If you like these kinds of analyses, you can find more AI and machine-learning research on AIModels.fyi or follow us on Twitter.
A benchmark tests whether agents can build the scene and make it readable
VibePrevisBench evaluates coding agents that turn a brief into an executable Blender scene and a graybox proxy video—the rough video used to guide a later video generator. Across 242 tasks and 11 agents, the best overall score is 77.7, and no agent reaches a Rule-based Scene Check score of 90 on more than 35 tasks. The benchmark’s central finding: agents can produce plausible previews, but still struggle to reconstruct demanding scenes and verify that their choices match the brief.
The benchmark targets a practical failure: errors in a proxy video can carry into the generated final video. It therefore checks both whether the underlying 4D scene is correct and whether viewers can recognize the intended subjects, actions, and story in the rendered video. For a related benchmark focused on finished code-generated videos rather than the 4D scene behind a proxy, see VibeVideoBench; it scores exact content and timing in playable MP4s.