Counting in vision-language models (VLMs) follows a peculiar pattern: exact on small sets, then drifting into estimation as the number of items grows. That profile, reported in a new arXiv paper, is not arbitrary. It resembles the performance of children before they have fully mastered counting.
Children typically cross this barrier through epistemic actions—physically pointing at each item as they count. The paper's title, "Give the Model a Finger," signals a developmentally aligned evaluation strategy: rather than treating counting as a purely perceptual task, the authors argue for evaluating VLMs in a way that mirrors how children learn to ground number words in one-to-one correspondence.
Because the source is a single preprint, there is no independent comparison to report. The contribution is a framing: VLM counting errors may be less a failure of arithmetic and more a missing sensorimotor scaffold. Whether giving models a "finger" actually closes the gap is the empirical question the paper opens.