How it works
Tying a claim to a region of the image
Grounded models tie each statement to a region, which makes a score checkable in a way a bare number never is.
Guides on How it works: The whole training run, in order, Classifier, regressor or language model: three scorers, three kinds of answer, Every score is a memory of somebody's opinion, An axis earns its place by being separately observable, What an image becomes before it is scored
Grounding means a vision-language model points at the region of the photo each claim is about, so a score can be checked instead of trusted. Most scores arrive with nothing attached: a number, maybe a sentence, and no way to check either against the photo. Grounding closes that gap by naming which pixels a conclusion rests on.
What grounding means technically
A grounded vision-language model can produce a bounding box, a segmentation mask, or a coordinate reference alongside a piece of text, tying that specific statement to that specific region of the image. Kosmos-2 (Peng and colleagues, 2023) is a well-known example: it writes phrases as links to bounding boxes, so the text and the location arrive together. Ask an ungrounded model to describe symmetry and it gives you a sentence. Ask a grounded one and it can give you the sentence plus a highlighted region the sentence is about, which is a fundamentally different kind of output - one you can check by looking at the same region yourself and asking whether it supports the claim.
This is distinct from attention weights inside a network, which show which input regions influenced a computation but were never designed to be a human-readable claim about content. Jain and Wallace's 2019 study of attention concluded that standard attention modules "do not provide meaningful explanations"; grounding is a deliberate output format instead, built so a statement and a location travel together. It is also a distinct concept from hallucinated description - a model can be grounded and still occasionally point at the wrong region, but the grounding is what makes that error visible in the first place, rather than hiding it behind confident prose with nothing to check it against.
Why it matters specifically for scoring
A bare score is a black box by construction: input goes in, number comes out, and nothing in between is inspectable. A written explanation without grounding is only marginally better, because prose is persuasive independent of whether it is accurate - a fluent paragraph reads as evidence even when it is not tied to anything checkable.
Grounding changes what a person can do with a result. If a tool states that a particular region drove the shape score down, and shows that region, a user can look at the actual photo and agree or disagree with the specific claim, rather than accepting or rejecting an unverifiable total. This is the difference between a result you can audit and a result you can only feel, and it is the same underlying idea that makes a per-axis breakdown more useful than a single number taken one level further: not just which axis moved, but which part of the photo moved it.
The limits of grounding
Grounding tells you where a claim points, not whether the claim is correct. A model can ground a statement confidently in a region that does not actually support it, and the pointing itself carries no guarantee of accuracy - it is a step toward auditability, not a substitute for it. It also does nothing for the parts of a score that are not about any single region at all, like an overall-impression axis that is closer to a holistic judgement than a localised one; those remain as unverifiable as they were before grounding existed.
Grounding is also unrelated to whether the model's stated reasoning actually produced its answer, a separate and harder problem about the faithfulness of generated text in general, not specific to spatial claims.
Where this sits in a real pipeline
General-purpose vision-language models now return boxes and coordinates on request, but providers hedge them; Anthropic's vision documentation calls its model's localisation outputs "approximate." Few consumer scoring tools expose grounding to users at all - it remains more common in research systems and specialised fields like medical imaging. Rate Cock moves in this direction with its per-axis chart, which at least narrows an unverifiable total down to six narrower, more checkable claims, short of full spatial grounding but built on the same instinct that a claim should point at something.
A physical measurement sidesteps the whole question, because there is nothing to ground when the number comes from a tape against the subject directly rather than from a model's interpretation of a photo. Penis Rater's tool coverage looks at which scoring tools show their work at all, grounded or not, which is the more immediate practical question for most users. A human reviewer grounds every claim by default, in the sense that a person writing "the issue is here" can point at the actual photo while saying it - grounding is not a feature a human reviewer needs to be built, only a habit they need to keep.