驾驶前先思考：受世界模型启发的多模态视觉定位
Think Before You Drive: World Model-Inspired Multimodal Grounding
将世界模型的未来潜状态推演引入视觉定位，用命令感知的潜变量序列为歧义指令提供前瞻线索，这一「先推演后定位」的范式可迁移到任何需要空间关系推理的视觉-语言任务。
原文依据 we propose ThinkDeeper, a framework that reasons about future spatial states before making grounding decisions
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论解读自然语言指令以定位目标物体，对自动驾驶（AD）至关重要。现有自动驾驶中的视觉定位（VG）方法难以处理歧义性强、依赖上下文的指令，因为它们缺乏对三维空间关系的推理以及对场景演化的预判。基于世界模型的思想，我们提出 ThinkDeeper，一个在做出定位决策之前先对未来空间状态进行推理的框架。其核心是空间感知世界模型（SA-WM），它通过将当前场景蒸馏为命令感知的潜变量状态，并展开一系列未来潜变量状态来学习前瞻性推理，从而为消歧提供前瞻线索。与之互补，一个超图引导的解码器随后将这些状态与多模态输入进行层次化融合，捕获高阶空间依赖以实现鲁棒定位。此外，我们提出 DrivePilot，一个自动驾驶中的多源视觉定位数据集，其语义标注由检索增强生成（RAG）与思维链（CoT）提示的大语言模型流水线生成。在六个基准上的实验结果表明，ThinkDeeper 在 DrivePilot、MoCAD、Talk2Car 以及 RefCOCO/+/g 上优于当前最优基线。值得注意的是，它在挑战性场景（长文本、多智能体、歧义）下还表现出很强的鲁棒性与效率，并且即使在仅使用 50% 数据训练时仍保持优越性能。
英文原文摘要 1404 字符
Interpreting natural-language commands to localize target objects is critical for autonomous driving (AD). Existing visual grounding (VG) methods in AD struggle with ambiguous, context-dependent instructions, as they lack reasoning over 3D spatial relations and anticipated scene evolution. Grounded in the principles of world models, we propose ThinkDeeper, a framework that reasons about future spatial states before making grounding decisions. At its core is a Spatial-Aware World Model (SA-WM) that learns to reason ahead by distilling the current scene into a command-aware latent state and rolling out a sequence of future latent states, providing forward-looking cues for disambiguation. Complementing this, a hypergraph-guided decoder then hierarchically fuses these states with the multimodal input, capturing higher-order spatial dependencies for robust localization. In addition, we present DrivePilot, a multi-source VG dataset in AD, featuring semantic annotations generated by a Retrieval-Augmented Generation (RAG) and Chain-of-Thought (CoT)-prompted LLM pipeline. Empirical results on six benchmarks show that ThinkDeeper outperforms SOTA baselines on DrivePilot, MoCAD, Talk2Car, and RefCOCO/+/g. Notably, it also exhibits strong robustness and efficiency in challenging scenarios (long-text, multi-agent, ambiguity) and retains superior performance even when trained on 50% of the data.