Giving vision models a pointing tool made their counting visible—not reliably better
Across nine models and 682 images, numbered marks exposed omissions, duplicates and phantom objects. The tool improved small-photo counts for only four models and produced a clearly positive interval for one.
The short version
- What happened
- The researchers tested nine vision-language models on 682 controlled arrays and photographs under seven tool and prompting conditions.
- Who it affects
- People using vision models to inventory objects, inspect images or support accessibility need more than a confident number. A visible trace can make mistakes easier to challenge, but the tested tool was not reliably accurate enough for consequential counting.
- What remains unproven
- This is an unreviewed workshop preprint; the developmental analogy is qualitative and does not mean that vision models learn or count like children.
Keep in mind
This is an unreviewed workshop preprint; the developmental analogy is qualitative and does not mean that vision models learn or count like children.
Models were accessed through one provider API, and the authors could not standardize reasoning settings for every endpoint.
The larger models ran only a stratified subset of conditions and images; the photo datasets cover limited object counts and categories.
No separate funding or competing-interest statement is reported; one author uses an Antigma.ai address.
Source analysis · 1 linked source. Primary analysed source date: 20 October 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
The marks turned a wrong answer into an inspectable failure
Four independent researchers gave vision-language models simple image tools: place a numbered tag, crop, erase or overlay a grid. The tools contained no visual intelligence. The model chose where to mark, then saw the updated image. Because the datasets supplied object locations, each tag could be classified as correct, duplicated, placed on nothing or missing an object.
The study covered nine models, 682 images and seven harness conditions. It combined 272 controlled dot arrays, 30 larger arrays, 200 FSC-147 photographs with 8–80 objects and 180 PixMo-Count photographs with 2–10. Three cheaper models ran every condition; four larger models were tested on a stratified 240-image subset, so not every model-condition comparison uses the full denominator.[1]
A visible procedure did not become a dependable skill
On small-count photographs, tags improved four of nine models and produced a paired-bootstrap interval excluding zero for one. Imprecise pointers became worse, and the tagging loop rarely beat a verbal counting routine or a list of unseen coordinates. Reversing the instructed scan order also changed answers for some models beyond ordinary repeat instability, suggesting that following a procedure did not guarantee order-independent counting.
The practical value is evaluation: a tester can see whether a model skipped objects, double counted or reported a number inconsistent with its own marks. The study used one API, temperature zero and mostly low reasoning effort; two endpoints could not disable reasoning, and the photo datasets cover bounded count ranges. Independent replication and training tests would be needed before calling this a path to reliable visual enumeration.[1]
The distinction that matters
What a visible counting trace reveals
A numbered mark can be checked against object-level ground truth.
On target
One mark lands on one object
Supports one-to-one counting
Duplicate or phantom
Extra mark repeats or misses an object
Shows over-counting
Omission
An object receives no mark
Shows where the count was lost