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

SpaceCast-Bench. Evaluating Predictive Spatial Reasoning in Vision-Language Models. By Hongxing Li and seven others, from Zhejiang University. Published on the eighth of October.

Most spatial reasoning benchmarks test perception. They ask a model to read off relations that are already visible in the input. Real spatial intelligence asks for more than that. It means building a scene from observations, anticipating how an intervention would change it, and reasoning about an outcome that nobody has shown you. The authors call this predictive spatial reasoning, and existing evaluation does not measure it directly.

SpaceCast-Bench is presented as the first benchmark to evaluate that capability directly and diagnostically. It is built around an observe, transform, infer framework: look at a scene, apply a change, then reason about the result.

There are three thousand eight hundred and sixty-two questions, drawn from one hundred and eighty-two real-world scenes. They span sixteen task types arranged in three levels. Static perception sits at the bottom. Local prediction comes next, then global prediction. Each level asks for more: first scene understanding, then updating the spatial state, and finally relational inference over outcomes that were never observed.

Twenty-one models were evaluated, and the gap is stark. The strongest model reaches only fifty-eight percent. Humans reach eighty-seven point two percent. Models built specifically for spatial tasks stay near random chance.

Controlled analyses add two observations. Bridge views turn out to be critical for integrating observations spread across a scene. And explicit 3D evidence helps models more reliably than generated images or videos of the outcome. Finally, fine-tuning on the benchmark's programmatically generated data lifts Qwen3-VL-4B from thirty-four percent to sixty-five point seven percent, with macro-average gains across six out-of-domain benchmarks.

A model that can describe what it sees has not necessarily learned to picture what happens next. SpaceCast-Bench separates those two abilities level by level, shows the best models well short of people on the second, and shows that generated training data can move a small model a long way toward it.