Evaluating 3D Creation Delivery and Spatial Reasoning in Large Language Models
This article abstracts a raw-prompt 3D creation experiment into a general evaluation problem for large language models: how should the quality of 3D creation and spatial reasoning deliverables be assessed when no additional skills are attached? Using a detective-agency-themed voxel diorama as a case study, it proposes a four-layer evaluation framework covering semantics, geometry, spatial relations, and delivery. It also presents a structured acceptance specification and pseudocode for scoring, analyzes the strengths and weaknesses of the delivered 3D model, renders, and browser interaction page, and outlines an engineering pipeline from prompt template and generation to acceptance, repair, and archiving for reproducible evaluation.
1. Background: 3D Creation as a New Observation Surface for Model Capabilities
For some time, capability evaluations for large language models have focused mainly on text understanding, code generation, and tool calling. With advances in multimodality and spatial reasoning, 3D creation is gradually becoming a new observation surface: a model must not only “describe” a scene, but also organize the objects, hierarchy, positional relationships, material styles, and delivery formats in that scene into inspectable assets. This type of task is more complex than pure code generation because it simultaneously tests semantic understanding, geometric construction, spatial layout, format delivery, and interaction packaging.
A recent new version of one type of large model made 3D creation one of its flagship capabilities and provided a delivery channel from modeling to a browser interaction page. This article does not promote or review a specific product. Instead, it abstracts the original experiment into a general technical question: under raw-prompt conditions, how can the quality of a large model’s 3D creation and spatial reasoning deliverables be evaluated? The original experiment used a “detective agency themed voxel diorama” as its case study. The deliverables included multi-angle rendered images, an editable 3D project file, and a browser interaction page. This case is representative: it has clear semantic anchors as well as requirements for spatial hierarchy and interactive presentation, making it suitable for building an evaluation framework.
It is important to emphasize that 3D creation evaluation cannot rely only on whether the result “looks right at first glance.” A deliverable 3D asset contains at least four layers: the semantic layer, the geometric layer, the spatial layer, and the delivery layer. The semantic layer checks whether the thematic elements are complete. The geometric layer checks volumes, topology, and editability. The spatial layer checks the relative positions of floors, staircases, entrances, characters, and other elements. The delivery layer checks whether the project file, rendered images, and interaction page are complete and usable. Only by separating these four layers can evaluation avoid being misled by a single rendered image or a flashy web page.
2. Evaluation Design: Raw Prompts and Acceptance Criteria
The original experiment used a raw prompt without attaching additional skills, which helps reveal the model’s actual performance under minimal context. The prompt required creating a detective-agency-themed voxel diorama in a voxel block language, using the agency as the visual highlight, and delivering an editable 3D project and rendered images. In actual delivery, the model produced two rendered images and one 3D model file. A browser page requirement was then added, and the page included drag rotation, auto-rotation, scene annotations, view focus, model download, and image switching.
From an evaluation perspective, acceptance criteria should be defined in advance for this kind of experiment. Otherwise, it is easy to develop a bias in which “a page with more features gets a higher score.” It is advisable to structure the prompt and explicitly specify required elements, spatial relationships, delivery formats, and interaction requirements. For example:
{"theme": "detective agency themed voxel diorama","style": "voxel","must_have": ["two-story agency", "side staircase", "ground-floor coffee shop", "facade sign", "entrance character"],"spatial_rules": ["the staircase connects the ground floor and second floor", "the agency is on the second floor", "the coffee shop is on the ground floor", "the sign is on the building facade"],"deliverables": ["editable 3D project file", "multi-angle rendered images", "browser interaction page"],"interactions": ["drag rotation", "auto-rotation", "scene annotations", "view focus", "model download"]}
During acceptance, pseudocode can be used to perform automated or semi-automated checks on the deliverables:
def evaluate(deliverables, spec):score = {}found = detect_semantic_elements(deliverables.model)score["must_have_recall"] = len(found & set(spec.must_have)) / len(spec.must_have)score["spatial_consistency"] = check_relations(deliverables.model, spec.spatial_rules)score["editability"] = inspect_editable_assets(deliverables.project)score["render_quality"] = assess_render(deliverables.images)score["interaction_completeness"] = check_interactions(deliverables.web_viewer, spec.interactions)return score
The key here is that “required-element recall” and “spatial consistency” should take priority over rendering aesthetics. Rendered images can be improved through post-processing, but a missing facade sign, an incorrectly placed staircase, or confused floor relationships indicate that the model did not truly understand the scene semantics.
3. Deliverable Analysis: Strengths and Weaknesses Coexist
The case results show that the model restored the main building structure with a certain level of completeness: the side staircase, second-floor agency, ground-floor coffee shop, and entrance character basically appeared, indicating that the model can decompose the theme into several sub-objects and organize them into a voxel scene. The post-processing of the rendered images was also relatively good and formed a consistent visual style. The browser interaction page exceeded the basic requirements, providing drag rotation, auto-rotation, scene annotations, view focus, model download, and multi-image switching. If used for smart visualization, sandbox display, or urban modeling demonstrations, these features do have practical value.
However, the weaknesses are equally clear. First, semantic anchors were missing. In a detective agency theme, the facade sign is a key element for identifying the theme, but the model did not present it in the scene, reducing thematic recognizability. Second, the character was over-abstracted. Although an entrance character appeared, its shape and hairstyle were too abstract, indicating that the model remains unstable at fine-grained semantic levels. Third, interaction features must be distinguished from model capability. A feature-rich page may come from the model’s autonomous generation, or it may come from generic scaffolding or template completion; evaluation should check whether the features were generated by the model according to the prompt, avoiding misattributing template capabilities to the model. Fourth, editability needs verification. Delivering a project file does not mean the asset is editable; object naming, hierarchy organization, material assignment, and modification cost should also be checked.
Therefore, a reasonable score should be split into “basic restoration score,” “spatial relationship score,” “editability score,” “interaction enhancement score,” and “semantic anchor penalty.” Interaction enhancement can add points, but it cannot offset theme-level problems such as a missing facade sign.
4. Engineering Recommendations: From Demos to Reusable Pipelines
To apply 3D creation capabilities to real projects, it is advisable to establish a “prompt template—generation—acceptance—repair—archiving” pipeline. The prompt template should explicitly list required elements and spatial rules. After generation, object names and bounding boxes should be automatically extracted to check floor heights, staircase connections, and facade elements. When acceptance fails, the missing items should be fed back to the model for a second repair round. Rendered images should retain multiple angles and wireframe views to make geometric structure easier to inspect. Project files should require standardized object naming, clear hierarchy, and replaceable materials. Browser pages should modularize interaction capabilities, for example by separately accepting rotation control, annotation system, focus system, and download entry.
The evaluation method should also be reproducible: fix the theme, prompt, and delivery format, run the process multiple times, and record the element recall rate, spatial error types, and repair rounds for each run. For scenarios with limited resource quotas, evaluation should prioritize high-value dimensions such as semantic anchors and spatial relationships, rather than pursuing complex animations or large-scale scenes from the outset. Only by moving evaluation from “looking at rendered images” to “structured acceptance” can 3D creation capabilities truly enter production workflows.
