Coordinates Help When Words Cannot Point

A detailed ablation of grounded visual reasoning found that emitting text coordinates improved counting but produced no significant advantage over nongrounded baselines across a wider set of spatial-reasoning and perception benchmarks. Synthetic tasks isolated the useful case: references that are visually precise but hard to name. The authors therefore introduce BrickVision, which requires such undescribable references, and a trace analysis that measures whether a model actually uses its proposed grounding. The result narrows the claim for coordinate-based reasoning instead of treating it as a universal visual upgrade.