Grounding Generative Verification in 3D Content

TL;DR

The paper introduces 3DVerifierBench, a benchmark for evaluating 3D content verification across 10 tasks, revealing a significant performance gap between humans and current models. It presents a scalable pipeline to generate 14,126 training examples with contradiction explanations and counterfactuals, enabling better detection of factual inconsistencies. Experiments show that verification-driven candidate selection and recursive refinement dramatically improve factual accuracy in 3D generation.

Why It Matters

Reliable verification is essential for trustworthy 3D content creation, especially in safety-critical applications like robotics and virtual design. This work provides tools and data to build systems that not only generate 3D content but also critically assess and improve it. It advances the field toward self-correcting 3D generation pipelines grounded in factual consistency.

Key Findings

• Human performance on 3D verification tasks exceeds the best model, GPT-6 Astra, by 21.6 percentage points, indicating a major gap in reasoning and explanation quality.
• A scalable pipeline using multi-view analysis, controlled 3D edits, and native annotations produces 14,126 high-quality training examples with detailed contradiction explanations and counterfactuals.
• Oracle-based candidate selection increases factual success from 26.7% to 90.0%, and human-guided recursive refinement achieves full requirement satisfaction in all tested cases.