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

Forum topic · 小凯 · 2026-10-11

Summary

SpaceCast-Bench is the first benchmark designed to directly and diagnostically evaluate predictive spatial reasoning in vision-language models. Unlike existing benchmarks that test spatial perception (reading relations already visible in the input), SpaceCast-Bench assesses whether models can construct scenes from observations, anticipate how interventions change them, and reason about unseen outcomes. Built on an observe-transform-infer framework, it contains 3,862 questions from 182 real-world scenes covering 16 task types across three levels: static perception, local prediction, and global prediction. Evaluation of 21 models reveals a large gap between AI and humans: the strongest model reaches only 58.0% versus 87.2% human performance, while spatially specialized models perform near random. Controlled analyses show that bridging views are critical for integrating scattered observations, and explicit 3D evidence helps models more reliably than generated images or videos. Fine-tuning on programmatically generated data lifts Qwen3-VL-4B from 34.0% to 65.7%, with average gains across six out-of-distribution benchmarks.

Paper Overview

Research Area: NLP Authors: Hongxing Li, Jinyue Su, Dingming Li, Wenqi Zhang, Weiming Lu, Jun Xiao, Yueting Zhuang, Yongliang Shen Published: 2026-10-08 arXiv: 2610.12402

Summary

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 — scene understanding
Local prediction — spatial state updating
Global prediction — relational inference over unobserved outcomes

Key Findings

Evaluating 21 models exposes a stark gap: the strongest model achieves only 58.0%, compared to 87.2% human performance.
Spatially specialized models perform nearly at random level.
Controlled analyses show that bridging views are crucial for integrating scattered observations.
Explicit 3D evidence helps models more reliably than generated result images or videos.
Fine-tuning on programmatically generated data improves Qwen3-VL-4B from 34.0% to 65.7%, with macro-average gains across six out-of-distribution benchmarks.

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

Summary

Paper Overview

Summary

Key Findings

Links

Tags