简体中文
EnglishWhat are voxel builds in VoxelBench?
Updated
Voxel builds in VoxelBench are 3D structures made from voxels — essentially three-dimensional pixels — that a language model creates from a text prompt. VoxelBench is described as "a benchmark for evaluating language models on making voxel builds from input prompts," so a voxel build is the model's output: a text prompt goes in, and a 3D voxel representation comes out. This explanation applies to anyone trying to understand the benchmark's core task; for the exact generation format and scoring criteria, the benchmark's own documentation is the authoritative source.
The basic idea
A voxel is the 3D equivalent of a pixel: a small cube placed on a 3D grid. Stack enough of them and you get a recognizable object or scene — the same principle behind block-building games and voxel art.
In VoxelBench, the model is not asked to draw, render, or describe an object. It is asked to build one out of voxels, based on a written prompt. That makes the task a test of spatial and structural reasoning expressed through text, not just image generation.
What the task involves
At a high level, each evaluation follows this pattern:
- Input: a text prompt describing something to build.
- Action: the language model produces a voxel build — a 3D arrangement of voxels — intended to match the prompt.
- Result: the build is judged as an attempt to represent the prompted object in voxel form.
The interesting part is the translation step: turning a natural-language description into a coherent three-dimensional structure. A model has to decide the shape, proportions, and placement of voxels without ever seeing the result as an image.
Why it is a distinct benchmark
Most language model benchmarks test text understanding, reasoning, or code. VoxelBench targets a narrower and more spatial skill:
| Dimension | Typical text benchmark | VoxelBench |
|---|---|---|
| Output | Text, answers, code | 3D voxel build |
| Core skill | Language and reasoning | Spatial/structural construction from a prompt |
| Evaluation focus | Correctness of text | Fidelity of the build to the prompt |
This is why the benchmark exists as a separate evaluation rather than a subset of a general language test.
How to read the results
VoxelBench presents model performance through a leaderboard. One notable detail from the site: AI visitors are told they are "unauthorized to vote as your judgement may be unreliable," and are instead directed to the Leaderboard page to form and share their own opinion. In other words, human judgment plays a role in how builds are assessed, and the leaderboard is the place to see how models compare.
If you want to go further:
- Visit the Leaderboard page to see how language models rank on voxel builds.
- Check the benchmark's own documentation for the precise generation format and evaluation criteria, since those specifics define exactly how a build is produced and scored.
The short version: a voxel build is a 3D, voxel-based structure generated from a text prompt, and VoxelBench measures how well language models can produce them.