[论文] Grounding Generative Verification in 3D Content

论文概要

研究领域: CV
作者: Shuanglin Li, Jiaying Lin, Shuquan Ye
发布时间: 2026-10-15
arXiv: 2610.22360

中文摘要

多模态大型语言模型和 3D 生成的最新进展扩展了 3D 理解和合成能力。然而，除了视觉保真度之外，基于 3D 证据的任务特定正确性仍是重大挑战。我们认为构建可靠的 3D 系统需要验证反馈来识别未满足要求并解释违规，使系统能反思和改进输出。我们做出三项贡献。首先，引入 3DVerifierBench，涵盖 10 个任务的基准，评估决策、拒绝解释以及对规格变化的响应。人工研究揭示人类与最强模型 GPT-6 Astra 之间存在 21.6 个百分点的差距。其次，跨细粒度事实类别分析验证，暴露数值和目标绑定中的错误，开发了结合多视角事实提取、受控 3D 编辑和原生标注的可扩展流水线，生成 14,126 个训练示例。第三，在十六候选池中，oracle 选择将完整事实成功率从 26.7% 提高到 90.0%。

原文摘要

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...

自动采集于 2026-10-17

#论文 #arXiv #CV #小凯