[论文] GeoThinker-F: Endogenous Predictive World Representations for Fast Spa...

论文概要

研究领域: CV
作者: Haoyuan Li, Qihang Cao, Huanyuan Zhou, Xin Hu, Sijia Chen
发布时间: 2026-10-14
arXiv: 2610.20592

中文摘要

现有的几何感知多模态大语言模型（MLLM）越来越依赖外部3D编码器提供结构先验，但在推理时执行这些额外的编码器路径会引入大量延迟和计算开销。我们提出GeoThinker-F，一个通过将因果MLLM隐藏状态转化为预测性潜在世界表示来内生化时空感知的高效框架。我们的核心洞察是：中间语义表示包含足够信息来预测当前几何和未来语义的有用表示，无需精确复现专用3D编码器的特征。为实现这一思想，GeoThinker-F采用解耦的两阶段训练策略。第一阶段冻结MLLM骨干，轻量级潜在帧预测（LFP）头从离线3D几何专家和未来帧语义目标中学习。第二阶段移除外部专家，模型通过专门的Spatial-Grounded Temporal（SGT）融合模块联合优化，生成并选择性整合自预测先验。大量评估表明，GeoThinker-F消除了对外部3D几何编码器的运行时依赖，同时在多样化空间智能基准上保持有竞争力的性能。它实现高达3.6倍的推理加速，将延迟从1.45秒降至0.40秒，同时保持强大的空间推理性能。

原文摘要

Existing geometry-aware Multimodal Large Language Models (MLLMs) increasingly rely on external 3D encoders to provide structural priors, but executing these additional encoder paths at inference introduces substantial latency and computational overhead. We propose GeoThinker-F, an efficient framework that internalizes spatiotemporal perception by transforming causal MLLM hidden states into predictive latent world representations. Our key insight is that intermediate semantic representations contain sufficient information to predict useful representations of current geometry and future semantics, without requiring an exact reproduction of features from a dedicated 3D encoder. To realize this idea, GeoThinker-F adopts a decoupled two-stage training strategy. In the first stage, the MLLM back...

自动采集于 2026-10-16

#论文 #arXiv #CV #小凯