Letting an AI model "think again" doesn't help if it can't tell when it got it right.
Researchers built a system called Iterative Visual Thinking (IVT) that asks a vision-language model to predict a bounding box around an object, render that box on the image, and then refine its guess over several steps. The training pipeline is clever: it uses the model's own wrong predictions as realistic errors, has a teacher model generate corrective reasoning traces from those errors, and then fine-tunes with a reward tied to how well the predicted box overlaps the target. Benchmarked the usual way, IVT looked like a win — a +2.4 percentage point accuracy gain over the single-shot baseline.
The catch is how that number was computed. The evaluation kept, per example, whichever step in the refinement trajectory came closest to the correct answer — a selection that requires knowing the ground truth. Under any stopping rule a real deployment could actually use, the gain disappears entirely. The best policy turns out to be stopping at step zero, which matches the base model and beats every label-free rule tested. The model can stumble onto a better box somewhere in its chain of attempts, but it has no reliable way to know when it has done so: self-verification confidence correlates with correctness at roughly r = 0.22, barely above noise.
This matters beyond one paper. "Test-time compute" — the idea that more inference steps produce better outputs — has become a central narrative in AI research, and spatial grounding is a natural place to apply it. IVT is a careful, well-constructed version of this idea, and it still fails the deployability test. The authors propose an evaluation protocol that reports accuracy under fixed label-free policies alongside an explicit oracle-to-shippable gap — a useful corrective for a field that has leaned hard on oracle metrics to sell progress.