SpaceCast-Bench: Evaluating Predictive Spatial Reasoning in Vision-Language Models
arXiv cs.CL · 9 Oct 2026, 03:44 AEST
SpaceCast-Bench is introduced as the first benchmark to evaluate predictive spatial reasoning in vision-language models. It highlights a significant performance gap between models and human capabilities in spatial reasoning tasks.
Key facts
- Existing spatial reasoning benchmarks mainly test spatial perception: reading off relations already visible in the input.
- We introduce SpaceCast-Bench, the first benchmark to directly and diagnostically evaluate this capability.
- 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.
Read aloud
SpaceCast-Bench is introduced as the first benchmark to evaluate predictive spatial reasoning in vision-language models. It highlights a significant performance gap between models and human capabilities in spatial reasoning tasks. Existing spatial reasoning benchmarks mainly test spatial perception: reading off relations already visible in the input. We introduce SpaceCast-Bench, the first benchmark to directly and diagnostically evaluate this capability. Evaluating twenty one models exposes a stark gap: the strongest model reaches only fifty eight point zero percent against eighty seven point two percent 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.
Sources in this cluster
- SpaceCast-Bench: Evaluating Predictive Spatial Reasoning in Vision-Language Models
arXiv cs.CL · 9 Oct 2026, 03:44 AEST