Vision-language models can describe a photo, answer questions about it, and read text embedded in the image, but ask them how many objects are in the frame and they often fail. A new arXiv preprint, published 13 July 2026, reveals something stranger: these models frequently encode the correct count internally, then output the wrong number anyway [S1]. A fix that detects those internal errors and selectively re-prompts the model lifts counting accuracy by up to 15.6 percentage points without any retraining [S1]. The finding raises a question anyone deploying AI for visual tasks should want answered: if the model knows the answer and still gets it wrong, what else is it silently misreading?

The count is there, but misaligned

Vision-language models — AI systems that process both images and text, like those powering visual search and image-based chatbots — have grown capable across many multimodal tasks. Counting remains a stubborn exception [S1].

Researchers from MBZUAI, DataBayt.AI Labs, and Imperial College London trained small diagnostic models, called probes, on the internal activations of four vision-language models across five counting datasets [S1][P2]. These probes are simple pattern detectors: they look at the model's internal state and try to predict whether the model is about to get the count right or wrong.

The probes worked. Nonlinear probes could reliably detect counting errors before the model produced its answer [S1]. More strikingly, probes trained on ground-truth counts and those trained on the model's actual outputs occupied a partially shared activation subspace — the same internal space, but read out along misaligned directions [S1]. In plain terms: the model has the right information. It just looks at it from the wrong angle.

The team confirmed this with a causal steering experiment, nudging the model's internal representations along the direction the correct-count probes identified. Counting performance improved, proving that misalignment, not missing knowledge, is the root cause [S1].

A selective fix, not a retrain

The practical output is a detector-guided self-correction method. Instead of re-prompting the model every time (slow and expensive) or retraining it (slow, expensive, and often impractical for deployed systems), the method runs a lightweight internal detector. When that detector predicts the model is about to fail, only then does it re-prompt [S1].

The result: up to 15.6 absolute percentage points of improvement in counting accuracy, with no parameter updates [S1]. "Up to" matters here. It is a maximum across the tested datasets and models, not a guaranteed average gain.

What it means

The core finding reframes how we should think about AI errors. When a vision-language model miscounts objects in an image, the problem is not that it failed to see them. The model saw them, represented the quantity correctly in its internal state, and then fumbled the readout. The knowledge exists. The wiring between knowledge and output is broken.

This has implications beyond counting. If models can encode correct information but fail to express it, the same pattern may appear in other tasks: spatial reasoning, attribute identification, relationship detection. The probe methodology offers a diagnostic tool. Before trusting a model's output, check whether its internal state agrees with what it says. The gap between internal representation and external output may be a general failure mode, not a counting-specific quirk.

The gap between capability and reliability is where the real risk lives. This study adds another dimension to that gap: the model may not even be unreliable in the way we assumed. It may be reliable internally and unreliable only at the output stage.

What it means for business

A two-person e-commerce startup using vision-language models to auto-generate product descriptions from photos already feels this problem. "Includes 6 items" when the image shows 4. "Set of 12" when it is 8. Every wrong count is a customer complaint, a return, or a listing correction.

The detector-guided method suggests a practical workflow shift. Instead of blindly trusting the model's count or manually verifying every output, a business could run a lightweight error detector alongside the model. When the detector flags low confidence, route that image to a human reviewer. When it passes, ship the output. This selective checkpoint catches errors where they are likely without slowing down the entire pipeline.

For a suburban real estate agency using AI to count rooms or features from property photos, the stakes are different but real. A model that says "3 bedrooms" when the photo shows 2 creates a compliance problem. The internal-detector approach means the agency could flag uncertain counts for manual review rather than verifying every listing by hand.

The cost profile matters. No retraining means no GPU time, no data labelling, no model deployment cycle. The probe runs on the model's existing activations. For small operators without ML engineering teams, that is the difference between a fix they can use and one they cannot.

What we don't know yet

The study tested four vision-language models across five counting datasets [S1]. Whether the findings generalise to other models, other tasks, or real-world image distributions remains open. The paper is a preprint and has not been peer-reviewed [S1].

The "up to 15.6 percentage points" figure is a maximum, not an average. Improvement varied across datasets and models, and the paper does not report a single headline number that represents typical performance.

The causal and mechanistic claims, including the SVCCA analysis and the steering intervention, rely on the authors' own interpretation of activation subspaces [S1]. Independent replication is needed before these become accepted explanations rather than one team's analysis.

A related line of work on text-only LLMs, the "Odometer Hypothesis" [P3], and a separate diagnostic framework for counting bias in VLMs called CounterCount [P4], suggest this problem is attracting broader research attention. Whether the detector-guided approach works on tasks beyond counting is the next question worth watching.

If your business relies on AI to count, measure, or quantify anything from images, the safest assumption right now is that the model may know more than it says, and you need a way to check.

Subscribe for more analysis on what AI actually does, not what the demo video showed.

Sources: [S1] arXiv preprint, "The Count Is There, but Misaligned: Understanding and Correcting Counting Failures in VLMs," 13 July 2026. [P2] Full HTML version, arxiv.org. [P3] GitHub repo, sohv/repeated-token-counting. [P4] arXiv preprint, "CounterCount: A Diagnostic Framework for Counting Bias in Vision Language Models."

Sources

More from Not A Tech Guy

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.