EVENT DOSSIER
On September 7, 2026, arXiv published the benchmark for visual language models (VLM) called MultihopSpatial. This research aims to address the issue that existing VLMs ignore multi-step combined reasoning and precise visual localization. The main contributions include: constructing a comprehensive multi-step combined spatial reasoning benchmark covering complex queries of one to three steps; proposing the Acc@50IoU metric for synchronously evaluating reasoning and visual localization capabilities; and releasing the dedicated large-scale training corpus for MultihopSpatial-Train. Extensive evaluations show that combined spatial reasoning remains a significant challenge, but by training the model through reinforcement learning on the corpus, its intrinsic spatial reasoning abilities and downstream embodied operation performance can be improved.
arXiv:2603.18892v2 发布 MultihopSpatial，旨在解决现有视觉语言模型（VLM）基准测试忽视多步组合推理与精确视觉定位的问题。该研究提供三项贡献：一是构建涵盖 1 至 3 步复杂查询的综合性多步组合空间推理基准；二是提出 Acc@50IoU 指标，同步评估推理与视觉定位能力；三是发布 MultihopSpatial-Train 专用大规模训练语料。对 37 个最先进 VLM 的广泛评估揭示了组合空间推理仍是严峻挑战，并证明在语料上进行强化学习后训练可提升模型内在空间推理能力及下游具身操作性能。