# SpaceCast-Bench

Spatial reasoning evaluation shifts from reading relations already visible in the input to predicting the unseen outcome of an intervention, and SpaceCast-Bench splits that ability into static perception, local prediction and global prediction across 3,862 questions from 182 real scenes, where the strongest model reaches only 58.0% against 87.2% for humans.

Spatial reasoning evaluation shifts from reading relations already visible in the input to predicting the unseen outcome of an intervention, and SpaceCast-Bench splits that ability into static perception, local prediction and global prediction across 3,862 questions from 182 real scenes, where the strongest model reaches only 58.0% against 87.2% for humans.