cs.CVOct 15, 2026

Grounding Generative Verification in 3D Content

Authors: Shuanglin Li, Jiaying Lin, Shuquan Ye

Organizations: Southeast University · Shenzhen Loop Area Institute · The Hong Kong University of Science and Technology

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Recent advances in multimodal large language models (MLLMs) and 3D generation have expanded 3D understanding and synthesis. However, beyond visual fidelity, task-specific correctness grounded in 3D evidence remains a major challenge. We argue that building reliable 3D systems requires not only stronger generation and reasoning, but also verification feedback that identifies unmet requirements and explains their violations, allowing systems to reflect on and refine their outputs. Such feedback must integrate multi-view observations and distinguish genuine inconsistencies from projection-dependent appearances. We make three contributions. First, we introduce 3DVerifierBench, a benchmark spanning 10 tasks that evaluates decisions, rejection explanations, and responses to controlled changes in specifications and candidates. A human study reveals a 21.6 percentage-point gap between humans and the strongest tested model, GPT-6 Astra, while explanation-aware analysis exposes weaknesses in rejection reliability. Second, we analyze verification across fine-grained factual categories, exposing errors in values and target binding. We develop a scalable pipeline combining multi-view fact extraction, controlled 3D editing, and native annotations to produce 14,126 training examples with contradiction explanations and linked counterfactuals. Third, we examine how verification could support 3D generation through candidate selection and recursive refinement. In human-labeled sixteen-candidate pools, oracle selection improves complete factual success from 26.7% to 90.0%. In a separate eight-input study, human-guided recursive refinement satisfies all seven requirements for every input, compared with three under parallel Oracle@8. Together, our benchmark, analysis, and data pipeline establish a foundation for models to reflect on 3D content through reliable verification feedback.

Figures & tables

Figure 1: Even state-of-the-art MLLMs, e.g. , GPT-6 Astra, still struggle with 3D content verification. Top: Astra mistakenly accepts an unnatural rightward collapse and an identity-changing face edit. Bottom: We found that Astra is still far behind human performance across all tasks in our 3DVerifierBench. Views and text are abbreviated for display.
Figure 2: Evaluation scope relative to prior work. Prior work ( Zhang et al., 2025a ; Zhang et al., 2025b ; Zhang et al., 2026b ; Duggal et al., 2025 ) emphasizes preference, hierarchical quality, general visual verification, or asset-level consistency and artifact localization. Our 3DVerifierBench targets complementary verifier-side questions through explicit requirements, controlled counterfactuals, heterogeneous 3D evidence, and verification-guided refinement.
Figure 3: Overview of 3DVerifierBench. Each task presents task-specific 3D evidence together with a verification requirement, reference answer, and explanation. Evidence spans multi-view objects, marked parts and regions, grasp poses, identity comparisons, photo-to-geometry comparisons, and motion sequences. The table summarizes the number of examples and input evidence for each task, totaling 2,581 examples. Mat., Ground., Aff. Reg., and FineGeo. denote Material, Part Grounding, Affordance Region, and Fine-grained Geometry, respectively. Views and text are abbreviated.
Figure 4: Our pipeline. Reviewed evidence establishes positive anchors. Spec-CF changes the requirement under fixed evidence, whereas Cand-CF changes the candidate under the original requirement. Native annotations and task-specific evidence complement these controlled mutations. Automated checks, model-assisted review, and human review together produce the benchmark.
Model / EvaluatorObjectPartMat.Ground.RelationAff. Reg.GraspIdentityFineGeo.MotionOverallExpOA
4-CLSBinary
Human Performance
Human95.082.590.092.595.090.080.092.980.070.080.087.8–
Baselines
Random guess50.050.050.050.050.050.050.050.050.025.050.050.0–
Proprietary MLLMs
Table 1: Results on 3DVerifierBench. All values represent the binary Balanced Accuracy, except for the 4-CLS column (four-class Exact Match accuracy). Bold marks the best model score.
MaterialObjectPartAff. Reg.GraspGround.RelationIdentityFineGeo.MotionOverall
93.390.093.383.386.786.7100.093.3100.083.391.0
Table 2: Human validation of the explanation Judge. We sample 30 correctly rejected model responses per task. The numerical value in the bottom row is the Agreement rate (%), which represents the ratio of identical Judge and human assessments of explanation validity.
Figure 5: Typical GPT-6 Astra failure cases and findings of limitations of MLLMs in 3D verification.
ModelObjectPartMaterialAvg.
Gemini-3.862.50 / 56.2568.75 / 18.7575.00 / 56.2568.75 / 43.75
Gemma-462.50 / 37.5062.50 / 50.0068.75 / 37.5064.58 / 41.67
Qwen3.568.75 / 50.0062.50 / 50.0062.50 / 50.0064.58 / 50.00
Qwen3.656.25 / 50.0068.75 / 50.0068.75 / 50.0064.58 / 50.00
Mean62.50 / 48.4465.63 / 42.1968.75 / 48.4465.63 / 46.35
Table 3: Evidence representation ablations. (a) Multi-view RGB v.s. raw 3D input. Each cell reports multi-view / raw OBJ Balanced Accuracy (%). (b) Motion verification under different evidence representations. Each cell reports binary / 4-class Acc.
SelectorFact Sat.FullStrict
Random80.826.716.6
GPT-6 Astra90.455.045.0
Human Oracle98.290.085.0
Table 4: What reliable verification could enable in 3D generation. Left: candidate selection from sixteen-candidate pools over 40 inputs. Right: human-guided recursive refinement on eight inputs. Full requires all seven facts; Strict additionally requires no severe defects and source fidelity.
Figure 6: Human-guided 3D recursive refinement trajectory. Given the source image and target facts, initial generation exhibits a missing side grille and a surface artifact (left, 6/7 facts). A revised one eliminates the artifact but leaves the grille incomplete (middle, 6/7 facts). Subsequent edit fully restores the mesh appearance and satisfies all facts (right, 7/7 facts). Intermediate 2D conditions and masks are omitted.
Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

TaskGPT-6 AstraGPT-5.6- SolGemini-3.8- FlashGemma-4- 31BQwen3.5- 122B-A10B
Object70.3%70.3%67.4%68.6%63.7%
Part67.2%62.3%57.6%50.7%53.4%
Material65.0%58.1%60.6%60.0%61.6%
Part Grounding86.2%81.1%79.0%64.4%58.3%
Part Relation88.0%78.1%84.9%69.8%66.1%
Affordance Region73.1%60.3%72.1%48.5%48.4%
Appendix
Table E3: Explanation-aware Balanced Accuracy by task (%). The two blocks show complementary sets of five models.
ConditionGPT-6 AstraGPT-5.6- SolGemini-3.8- FlashGemma-4- 31BQwen3.5- 122B-A10B
Type 150.0%50.0%52.0%52.0%50.0%
Type 256.0%50.0%52.0%50.0%54.0%
Type 350.0%50.0%50.0%52.0%50.0%
Type 466.0%52.0%68.0%72.0%56.0%
Appendix
Table E5: Identity condition-wise ExpBA (%). The four conditions stratify examples by expression and texture nuisance settings.
Fine FactUnitsGPT-6 AstraGPT-5.6- SolGemini-3.8- FlashGemma-4- 31BQwen3.5- 122B-A10B
identity_existence5978.0%64.4%72.0%66.9%65.3%
identity_preserving20055.5%50.5%55.5%56.5%52.5%
quantity12467.7%65.3%59.3%53.6%47.2%
shape_geometry26565.5%62.5%61.3%55.3%58.1%
spatial_configuration9483.5%72.3%78.7%64.4%58.5%
connectivity_contact10387.4%82.5%84.0%76.2%76.7%
Appendix
Table H3: Model-wise Fine Fact ExpBA (%). The two blocks show complementary sets of five models.
More in this topic

3D Scene Understanding

3D scene understanding builds geometric and semantic representations of environments for spatial reasoning, perception, and interaction.

See all 284 papers

Explore similar work

CardsList
Jul 12, 2026

3D-DefectBench: A Controlled Factorial Study of Vision-Language Model Evaluation Pipelines for Fine-Grained 3D Generation Defects

Automated evaluation is essential for scaling generative 3D systems, where exhaustive human review is costly and slow. Yet the reliability of an automated judge depends on the full evaluation pipeline, including the vision-language model (VLM), asset rendering, visual evidence, task specification, and human reference labels. We introduce 3D-DefectBench, a large-scale benchmark for rigorous evaluation-pipeline analysis. It complements holistic ratings and pairwise preferences with nine fine-grained binary defects spanning geometry, texture, and prompt adherence, with optional human severity annotations. Using a balanced factorial design, we vary the VLM, camera protocol, visual input, and prompt schema across 84 inference designs, and validate the resulting conclusions on a broader set of frontier models. Model choice is the dominant source of variation in agreement with human labels, while other pipeline factors also influence agreement, interact with the model, and can alter the best configuration. A compact six-view RGB protocol performs comparably to denser view sets and configurations augmented with depth or normal channels, making it a strong cost-effective default. Under this fixed design, the best of 12 VLMs still trail trained human labelers, and texture agreement drops sharply from expert-agreement to noisier silver labels. Severity annotations further show that binary judges recover most defects humans flag as severe. These results highlight the importance of evaluating automated judges as complete pipelines and calibrating them across human reference regimes.

Zhenyu Zhao, Nanshan Jia, Jihyeon Je +7

Roblox Corporation · Berkeley AI Research Lab & Department of Industrial Engineering and Operations Research, University of California, Berkeley · Computer Science Department, Stanford University +1

VLM Evaluation
AskPDF
Jun 9, 2026

P3D-Bench: Benchmarking MLLMs for Parametric 3D Generation and Structural Reasoning

Multimodal large language models can write code to produce complex programs as well as use programs to do 3D modeling, which opens up a new avenue for 3D generation powered by their priors, world knowledge and reasoning. Yet existing benchmarks rarely evaluate 3D modeling through code. Such modeling demands more than runnable code: from a text or visual specification, a model must generate a parametric 3D program that is geometrically precise, semantically aligned and assembly-consistent. We introduce P3D-Bench, a benchmark for parametric 3D generation. Unlike a 3D mesh, a parametric 3D program exposes explicit dimensions, construction operations and part relations, revealing whether a model recovers a design's structure, not just its appearance. Under a unified protocol, P3D-Bench covers three task families (Text-to-3D, Image-to-3D and Assembly-3D) and scores each output for executability, geometric fidelity, topology, text-grounded constraints, multiview semantic alignment and part-level structure. We construct P3D-Dataset, comprising 400 text cases, 400 image cases, and 203 annotated assemblies. Our evaluation on 100 cases from each task family yields three key findings. First, multi-part generation is substantially more challenging than single-part modeling, with models struggling to compose individual parts into a coherent structure. Second, models can often recover the global shape and semantic identity of the target object, yet fail to reproduce the precise parametric geometry specified by the input. Third, part-level modeling remains weak on assemblies, where models recover neither the geometry of each part nor the right number of parts. These results position P3D-Bench as a benchmark for evaluating precise parametric geometry and part-level structure in parametric 3D generation.

Yikang Yang, Zhanpeng Hu, Youtian Lin +5

Nanjing University · Envision

Part-Level 3D GenerationParametric CAD ModelingLLM-Based Program Synthesis3D Shape GenerationText-to-3D Generation
AskPDF
Oct 14, 2026

BlenderWorld: Benchmarking Visual Agents for 3D Creation from Real-World References

Recent advances in multimodal large language models (MLLMs) have enabled them to act as autonomous agents, creating and editing 3D models via code or direct software manipulation. In practical workflows, these models must infer object shapes, materials, spatial layouts, and lighting from diverse references while strictly following instructions to reconstruct, modify, or design 3D content. However, existing benchmarks provide limited coverage of such open-ended requests across real-world scenarios. To bridge this gap, we introduce BlenderWorld, the first comprehensive benchmark dedicated to evaluating frontier models on realistic 3D creation tasks. BlenderWorld comprises 223 tasks spanning architecture, products, natural environments, and scientific visualization, driven by text, image, video, and structured specification inputs. To rigorously assess these complex deliverables, we establish a unified evaluation protocol: we render the submitted Blender projects from multiple viewpoints and evaluate them against 2,651 task-specific binary rubrics alongside holistic visual quality scores. Benchmarking thirteen frontier models reveals that GPT-6-Astra achieves the highest overall fused score of 79.9, followed by Opus-5 (73.0) and GPT-5.6-Sol (67.7). Ultimately, our findings reveal a critical bottleneck: while current models can parse broad instructions, they still profoundly struggle with fine-grained 3D spatial coherence and accurate structural reproduction in Blender, leaving substantial headroom for future 3D-native agentics.

Xinyu Fang, Yuchi Wang, Weihao Yu +12

Zhejiang University · Alibaba Token Hub, Alibaba Group · The Chinese University of Hong Kong +2

3D Scene Generation3D Reconstruction Benchmarks
AskPDF