MJev-Spatial: System One Spatial Intelligence via Native Attention Matching

📄 arXiv: 2610.22276v1 📥 PDF

作者: Luozheng Qin, Yuhan Zhu, Changhai Zhou, Qian Qiao, Jun Gao, Yuhua Zhou, Harry Zhang, Xiangyu Zeng, Yuxuan Liu, Kaijie Chen, Andrew Chen, Pony Ma

分类: cs.CV

发布日期: 2026-10-15

备注: 37 pages, 7 figures, 27 tables. Code: https://github.com/Fr0zenCrane/MJev-Spatial

💡 一句话要点

提出MJev-Spatial以解决空间智能决策问题

🎯 匹配领域: 支柱九：具身大模型 (Embodied Foundation Models)

关键词: 空间智能 多模态决策 视觉-语言模型 分类框架 效率提升

📋 核心要点

  1. 现有视觉-语言模型在空间智能任务中效率低下，尤其是在自回归解码和答案解析方面。
  2. MJev-Spatial通过将每个候选项视为类别，利用预训练的查询和键投影进行评分，避免了自回归解码的复杂性。
  3. 该模型在多个基准测试中表现优异，训练速度提高30.8%，并在所有测试模型中取得最高得分。

📝 摘要（中文）

空间智能使代理能够在物理世界中感知和行动，许多判断需在交互速率下完成。现有的视觉-语言模型通过自由文本生成来回答空间关系、距离和位置等问题，然而这种自回归解码方式效率低下。为此，本文提出MJev-Spatial，一个统一的多模态决策模型，能够在一个分类框架内处理所有三种空间理解任务。MJev-Spatial将每个候选项视为一个类别，通过预训练的查询和键投影进行评分，且不增加参数。该模型在12.3M问答对上训练，取得了65.64%的最高得分，并在所有基准测试中表现优于自回归基线。

🔬 方法详解

问题定义：本文旨在解决空间智能任务中的决策效率问题，现有方法在处理空间关系、距离和位置时依赖自回归解码，导致效率低下。

核心思路：MJev-Spatial通过将每个候选项视为一个类别，并使用预训练的查询和键投影进行评分，避免了复杂的自回归解码过程，从而提高了决策效率。

技术框架：MJev-Spatial的整体架构包括候选项分类、粗到细的分类过程，以及联合监督训练。模型在推理时仅需一次前向传播，无需生成答案标记。

关键创新：该模型的主要创新在于采用统一的分类框架处理空间理解任务，显著提高了训练和推理的效率，与传统自回归模型相比，减少了计算复杂度。

关键设计：在训练过程中，采用了教师强制策略进行联合监督，并结合长度感知的全局批处理和跨GPU负载均衡，提升了训练吞吐量。

🖼️ 关键图片

fig_0
fig_1
fig_2

📊 实验亮点

MJev-Spatial在12.3M问答对上训练，取得了65.64%的最高得分，超越了八个测试模型中的七个基准，并在所有十二个基准测试中均优于未修改的自回归基线。该模型的推理速度比最快的自回归基线快2.5倍，且比基础模型快5.6倍。

🎯 应用场景

MJev-Spatial可广泛应用于机器人导航、自动驾驶、增强现实等领域，帮助系统更高效地理解和处理空间信息。其高效的决策能力将推动智能代理在复杂环境中的应用，提升用户体验和系统性能。

📄 摘要（原文）

Spatial intelligence lets an agent perceive and act in the physical world, and many of its judgments must be made at interaction rates: which spatial relation holds, how far an object is, or where to place it. Vision-language models (VLMs) produce these answers by free-form text generation, paying for autoregressive decoding and answer parsing even though the output space is compact: one of a few options, a scalar, or a 2D point. In light of this, we introduce MJev-Spatial, a System 1 multimodal decision model that tackles all three types of spatial understanding tasks within a unified classification framework. Specifically, MJev-Spatial treats each candidate as a class and scores it using the backbone's pretrained query and key projections, without adding parameters. The candidates are the given options for choice tasks, or value bins and image regions refined through coarse-to-fine classification for numerical and point prediction, respectively. At inference, each classification decision takes a single forward pass and generates no answer tokens. To train on heterogeneous spatial data efficiently, we supervise all coarse-to-fine stages jointly with teacher forcing in one forward pass, and combine this with length-aware global batching and cross-GPU load balancing, raising training throughput by 30.8% over the original batching strategy. Built on Qwen3.5-4B and trained on 12.3M question-answer pairs converted from existing spatial datasets, MJev-Spatial achieves the highest overall score (65.64%) among eight tested models, ranks first on seven of twelve benchmarks, and outperforms its unmodified autoregressive base model on all twelve, while running 2.5 times faster than the fastest autoregressive baseline and 5.6 times faster than the base model.