SpaceCast-Bench: Evaluating Predictive Spatial Reasoning in Vision-Language Models

Hongxing Li1*, Jinyue Su1*, Dingming Li1, Wenqi Zhang1, Weiming Lu1, Jun Xiao1, Yueting Zhuang1, Yongliang Shen1†
1Zhejiang University
*Equal Contribution, †Corresponding Author
arXiv Code 🤗Dataset

We introduce SpaceCast-Bench, a benchmark for predictive spatial reasoning: observing a scene, anticipating how an intervention changes it, and reasoning about the unseen outcome. It comprises 3,862 questions from 182 real-world scenes, spanning 16 task types across three capability levels: static perception, local prediction, and global prediction.

Abstract

Existing spatial reasoning benchmarks mainly test spatial perception: reading off relations already visible in the input. Yet real-world spatial intelligence demands predictive spatial reasoning: constructing a scene from observations, anticipating how an intervention changes it, and reasoning about the unseen outcome. We introduce SpaceCast-Bench, the first benchmark to directly and diagnostically evaluate this capability. Built around an observe–transform–infer framework, its 3,862 questions from 182 real-world scenes span 16 task types at three levels: static perception, local prediction, and global prediction, progressively requiring scene understanding, spatial state updating, and relational inference over unobserved outcomes. Evaluating 21 models exposes a stark gap: the strongest model reaches only 58.0% against 87.2% human performance, while spatially specialized models remain near random chance. Controlled analyses further reveal that bridge views are critical for integrating distributed observations, and that explicit 3D evidence benefits models more reliably than generated outcome images or videos. Fine-tuning on our programmatically generated data lifts Qwen3-VL-4B from 34.0% to 65.7% with macro-average gains across six out-of-domain benchmarks.

SpaceCast-Bench construction pipeline

Built around an observe–transform–infer framework, our geometry-grounded pipeline connects observations through depth-verified view chains, applies controlled 3D transformations, and derives labels programmatically. Models see only the initial scene and must infer spatial relations in the unobserved outcome.

Tasks span three capability levels: L1 (Static Perception), L2 (Local Prediction), and L3 (Global Prediction). They cover direction, distance, visibility and occlusion, and attachment structure, with directional relations posed in camera-centric, object-centric, and world-centric reference frames.

Task Types and Examples

Representative questions across static perception, local prediction, and global prediction. Green text marks the correct answers. Click the figure to inspect the scene images and question details.

Evaluation Performance

Evaluation of 21 models reveals a substantial gap: the strongest evaluated baseline reaches 58.0%, compared with 87.2% human performance, while spatially specialized models remain near random chance. Overall scores are macro-averaged across the 16 task types. These are the results reported in the paper.

Learning Predictive Spatial Reasoning

Staged supervised fine-tuning on our programmatically generated data raises Qwen3-VL-4B from 34.0% to 65.7% on SpaceCast-Bench and improves its macro-average score across six out-of-domain benchmarks from 33.5% to 38.1%. These results show that benchmark-aligned training supports both spatial state updating and transferable model improvement.

Citation

If you find SpaceCast-Bench useful, please consider citing our work:

@misc{li2026spacecastbenchevaluatingpredictivespatial,
      title={SpaceCast-Bench: Evaluating Predictive Spatial Reasoning in Vision-Language Models},
      author={Hongxing Li and Jinyue Su and Dingming Li and Wenqi Zhang and Weiming Lu and Jun Xiao and Yueting Zhuang and Yongliang Shen},
      year={2026},
      eprint={2610.12402},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.12402},
}