Evaluating General-Purpose LLMs on 3D Creation and Spatial Reasoning: A Practical Benchmark

    This article presents a hands-on evaluation of how a general-purpose large language model handles a voxel-style miniature diorama task. By testing semantic understanding, spatial reasoning, structured asset organization, code generation, and interactive web delivery, it identifies both strengths and weaknesses of current models and offers a reusable prompt template, a seven-dimension quality framework, and engineering practices for turning 3D generation into a verifiable production workflow.

    Why 3D Creation Is a New Lens for Evaluating LLMs

    As code generation becomes a baseline capability for general-purpose large language models, evaluation focus is shifting toward more complex composite tasks. 3D creation is a representative scenario: it simultaneously tests semantic understanding, spatial reasoning, structured asset organization, code generation, rendering, and interaction design. A model must not only understand a theme, but also decompose it into modelable elements, express them in a consistent geometric language, and finally deliver an editable project file, rendered images, and a browsable page. Compared with single-turn Q&A or function generation, this type of task is closer to real-world multimodal, multi-deliverable collaboration.

    In this evaluation, a general-purpose LLM was asked to build a voxel miniature diorama themed around a detective agency. Voxel language, with its regular grid, blocky structure, and clear spatial hierarchy, makes it easy to observe whether the model can consistently handle building volumes, floor relationships, stair connections, signage text, and character props. No specialized skills were attached; the test used raw prompts to observe the model’s native understanding and delivery capability.

    Evaluation Design and Prompt Strategy

    The evaluation used a “fixed theme + explicit deliverables + iterative supplementary requirements” approach. The first-round prompt asked the model to render the detective agency theme in voxel blocks, highlighting the main building, and to deliver an editable 3D project and rendered images. The second round added a requirement to generate an interactive page that opens in a browser. No skill templates were provided throughout; deliverable scope was constrained purely through natural language.

    A reusable prompt template is as follows:

    1. Task: Generate a voxel-style 3D miniature diorama.
    2. Theme: A detective agency, including a two-story main building, a side staircase, a ground-floor coffee shop, and character figures at the entrance.
    3. Style: Unified voxel block language, coordinated proportions, prominent main structure.
    4. Required elements: Building sign text, floor relationships, stair connections, two character figures.
    5. Deliverables: An editable 3D project file, several rendered images, and an interactive page viewable in a browser.
    6. Constraints: Clear project structure, identifiable element naming, and basic browsing operations on the page.
    7. Acceptance: Item-by-item check of required elements, spatial relationships, material consistency, and interaction usability.

    The value of this template is that it decomposes the “theme” into verifiable items. Otherwise, the model tends to produce results that “look like a building” while omitting the semantic elements most critical to theme recognition.

    Deliverables and Capability Observations

    In this sample, the model delivered two rendered images and an editable 3D project file. The main structure was essentially complete: a two-story agency, a side staircase, a ground-floor coffee shop, and two character figures at the entrance. This shows the model has some capability in spatial layout and volume organization, converting floor, entrance, and stair relationships from natural language into editable structures.

    However, the result also exposed a key problem: the sign text—the most prominent element of the theme—was missing. The sign is a strong semantic anchor for the “detective agency” theme; without it, theme recognition drops noticeably even when the building structure is similar. This indicates that explicit generation of salient semantic elements remains unstable, especially for objects combining geometry and text such as signs, placards, and entrance headers. These need to be explicitly emphasized in the prompt, ideally requiring them to be named and placed as independent objects.

    Character modeling also showed the abstraction typical of voxelization, with proportions and hairstyle details lacking accuracy. This is not purely an art issue; it reflects the trade-off between spatial reasoning and semantic detail. The model prioritizes overall scene coherence and applies weaker constraints to local character details. For engineering applications, acceptance checklists should mark characters, signs, entrances, and other key objects as “required items” rather than relying on the model’s free improvisation.

    Interactive Web Page: Over-Delivery Capability and Risks

    Notably, the prompt only asked for a file that opens in a browser, yet the model over-delivered features including model drag, auto-rotation, scene annotations, view focusing, asset download, image switching, and a first-open guide. For scenarios such as smart visualization, real-estate sand tables, and city modeling, scene annotation and view focusing have real value: users can quickly locate building components, lowering the comprehension cost of 3D scenes.

    This “over-delivery” also deserves an engineering review. First, more features mean a more complex page structure and higher maintenance and secondary development costs. Second, the model may produce interaction logic that looks complete but lacks configurability. Third, annotation labels must match model object naming; otherwise you get the “can focus but semantically wrong” problem. Therefore, interactive pages should be part of acceptance testing: verify object naming, annotation binding, view switching, asset paths, mobile adaptation, and graceful degradation.

    Quality Evaluation Dimensions

    Seven dimensions are recommended for evaluating 3D generation tasks:

    1. Semantic fidelity: Are key theme elements complete and accurate—sign, entrance, floor relationships?
    2. Spatial layout: Do building volumes, stair connections, and character placements follow common sense?
    3. Structural editability: Is the project file layered, clearly named, and are objects independently adjustable?
    4. Rendering quality: Do composition, materials, lighting, and stylistic consistency meet the bar?
    5. Interaction completeness: Are drag, rotation, annotation, focus, and download features functional?
    6. Robustness: Do supplementary requirements break the existing structure, and are asset references stable?
    7. Cost and quota: Are generation rounds, API quota consumption, and waiting time acceptable?

    Engineering Recommendations

    In real projects, 3D generation should not be treated as a one-shot black-box deliverable. A safer approach is staged delivery: first generate the main structure and key elements, then add materials and rendering, and finally build the interactive page—with a checklist-based acceptance at each stage. For key elements, use “explicit listing + object naming + position constraints,” for example requiring the sign to be an independent object placed above the entrance. For interactive pages, require an object inventory and annotation mapping to enable automated checking.

    A pseudocode acceptance example:

    1. checklist = ["two-story main body", "side staircase", "ground-floor coffee shop", "sign text", "two character figures"]
    2. for item in checklist:
    3. assert item in scene_objects, f"Missing key element: {item}"
    4. assert floor_relation["second floor"] > floor_relation["ground floor"]
    5. assert page.supports("drag", "rotate", "focus", "download")

    In addition, keep version control and human review. The model-generated 3D project and web code can serve as a first draft, but key semantic elements, brand identifiers, text content, and interaction bindings still need manual confirmation. When quota is limited, prioritize the call budget for “structure generation” and “key element correction” instead of repeatedly retrying low-value details.

    Conclusion

    This sample shows that general-purpose LLMs can already handle basic spatial layout, editable project delivery, and interactive page completion in voxel diorama tasks, demonstrating real potential for 3D creation and spatial reasoning. But the shortcomings are equally clear: salient semantic elements are easily omitted, local details lack stability, and interactive features carry a risk of over-delivery. For developers, the more reliable path is to integrate 3D generation into a verifiable engineering workflow—using structured prompts, key-element checklists, staged generation, and automated checks to constrain output. Only then can model capability be converted into reusable, maintainable productivity rather than a one-off demo effect.