Atom-Targeted Curation Against Geographic Hallucination

Vision-language models, which read an image and write text about it, invent details more often when the image comes from a low-income home in Nairobi than from a showroom in Munich. Two facts are already established. Vision models trained on web photos collapse on everyday objects outside the wealthy West. And a few hundred carefully written training captions can noticeably reduce invented details. What nobody knows is which few hundred images to pick when the goal is to close the geographic gap. Commissioning new photos costs about $0.87 each (GeoDE), so the budget question is the practical one.

We propose to split each model description into atomic claims, meaning single checkable statements such as "there is a stove," "the stove is metal," or "this is a Chinese dumpling." We then locate where false claims concentrate across region, household income, and claim type. The small fine-tuning budget goes to exactly those cells, and each chosen image gets a caption rewritten to keep the verified claims and correct the refuted ones. The two-sentence version is this. Claim-level error maps tell you both which images to buy or label and what their captions should say. We test whether that beats simply spending the same budget on a geographically balanced random sample.

Core Claim

At a fixed budget of 500 to 2,000 images, fine-tuning on images selected and recaptioned by their atomic-claim error profile will reduce the regional gap in claim-level factual precision more than two alternatives: a geographically balanced random subset with the same rewriting, and a subset selected by a single whole-response hallucination score. We also expect the gain to concentrate in the claim types that were targeted, and we treat that as a check on the mechanism. The idea stands or falls on the random-balanced comparison. If that comparison ties, the lesson is that geographic coverage, not claim targeting, is what matters. We would report that outcome as such.

Where This Sits

Three strands of prior work each supply part of this and leave part open.

Atomic decomposition gives the measuring instrument. FActScore splits long text into atomic facts and checks each against a source. It found a rarity penalty: precision drops for long-tail entities. Multi-FAct showed that the same penalty tracks region for text-only models. Both motivate our measurement, and neither touches images or training data.

In the multimodal setting, Pelican turns a response into predicates and checks them with code against a table built from off-the-shelf object detectors. ZINA labels hallucinated spans in six types: object, attribute, number, text, relation, and fact. We reuse ZINA's taxonomy and add one type, cultural attribution. Pelican also acts as a constraint on our design. Its detectors are exactly the kind of model that Dollar Street and GeoDE show failing on low-income and African imagery. A detector-based checker would therefore mislabel true regional claims as hallucinations, and the error map would be wrong precisely where it matters.

The data-quality literature supplies the premise that small budgets work. ReCaption fine-tunes on 500 images with ChatGPT-rewritten dense captions. Its rewrites are generated from extracted keywords with no grounding, and it evaluates on COCO-derived probes. SHED selects 1–10% of instruction data by approximate Shapley utility. GeoDE shows that adding a small diverse set improves Dollar Street by 9% and GeoDE by 21%. That last result is the strongest threat to our idea, because diversity alone might be enough.

The geographic-bias literature supplies the failure we target. When Tom Eats Kimchi (MIXCUBE) shows that multimodal models name cultural items from the ethnicity of the person holding them. It scores label correctness only, so it cannot say whether the model misperceived the object or overrode a correct perception with a prior.

Research Program

Diagnostic layer. The target model is LLaVA-1.5 at 7B and 13B, chosen because Volcano and MHR report baselines on it. The model writes detailed descriptions for a pool of GeoDE and Dollar Street images. We split both datasets by household or collection location so that near-duplicate photos from one home never span train and test. A language model decomposes each description into atoms, resolves pronouns so each atom stands alone, and tags each atom with one of the seven types.

Verification without detectors. We check atoms in two tiers, neither of which relies on a Western-trained detector. The first tier covers metadata-anchored atoms. These are claims about the object class, which both datasets label, and cultural-origin claims that contradict the known region. They are checked deterministically against metadata. The second tier covers remaining atoms such as attributes and relations. A strong vision-language verifier checks them with the image and the true object label in its prompt.

Because the verifier can itself be geographically biased, we audit it. Humans label a region-stratified sample of atoms, and we report verifier agreement per region and income quartile. That audit is also a finding in its own right, addressing Gap 2 in our survey. On the same sample we run a Pelican-style detector check, so the audit directly measures how much detector-based verification degrades by region.

Selection. The error map assigns each cell (region × income quartile × claim type × object class) a count of refuted atoms. Selection picks $B$ images to cover the highest-error cells, with diminishing returns per cell:

$$S^* = \arg\max_{|S|=B} \sum_{c} e_c \cdot \min\big(n_c(S),\, k\big),$$

Here $e_c$ is the refuted-atom rate in cell $c$ on the training pool, $n_c(S)$ is how many selected images contribute refuted atoms to $c$, and $k$ caps how many images any one cell can absorb. A greedy solution suffices because the objective is monotone and submodular.

Grounded rewriting. For each selected image, the new caption keeps verified atoms and replaces refuted ones with the metadata-correct statement. For example, "a religious building" becomes "a house." Unverifiable atoms are dropped. The result is then expanded into several dense, varied captions, following ReCaption's multi-rewrite recipe but restricted to verified content. We then fine-tune the model with the same recipe and budget ReCaption used.

Baselines

Base model. Unmodified LLaVA-1.5 is the do-nothing reference. On MMHal-Bench it scores 2.54 at 13B, and Volcano 13B reaches 2.64 there, with 88.3% overall POPE accuracy (Volcano). For LLaVA-1.5, or any vision-language model, we found no reported result on Lacuna for claim-level precision on GeoDE or Dollar Street. We establish that number ourselves as the first result of the diagnostic layer.

Random-balanced rewriting. This is the strongest competitor. It samples $B$ images uniformly across regions and income quartiles from the same pool and applies the same grounded rewriting, which isolates selection as the only difference. It is strongest because GeoDE's 21% gain from simple diverse data and ReCaption's gains from 500 random COCO images (7% on POPE and 10% F1 on FGHE for MiniGPT-4) both suggest that coverage plus good captions may be all that matters.

ReCaption as published. We apply ReCaption's ungrounded keyword-based rewriting to the random-balanced subset. This measures what grounding the rewrite contributes.

Full pool. We train on the entire training pool with metadata-derived captions. This is the quantity reference: is a few hundred targeted images competitive with tens of thousands?

Whole-response selection. This ablation isolates the key ingredient. It selects $B$ images by a single holistic hallucination score from the same verifier, with no decomposition, then rewrites identically.

Inference-time references. Volcano's self-revision and Pelican's detector-based correction are run on the base model, with latency reported alongside accuracy.

Fairness. All trained variants use the same base checkpoint, the same number of optimizer steps, three seeds, and the same small hyperparameter grid tuned on a held-out validation split from the pool. We claim a win only if the atom-targeted variant beats random-balanced on the regional precision gap by more than the seed spread at two or more budgets.

Why not just…?

Why not just fine-tune on a geographically balanced random sample? This is a real threat, and nothing in the literature rules it out. GeoDE's diverse augmentation already closes much of the classification gap. The matched-budget curves at 500, 1,000, and 2,000 images settle it directly. If random-balanced matches atom-targeted selection, we conclude that targeting adds nothing beyond coverage and say so. Our prediction of an advantage rests on errors being uneven across claim types, which the diagnostic layer measures before any training.

Why not just ask GPT-4o whether a description is hallucinated and select on that? On fine-grained multimodal hallucination detection, GPT-4o reaches 29.37 F1 against ZINA's 45.15 (ZINA). MIXCUBE finds GPT-4o swinging by up to 58% on low-resource cultures, so a holistic judge is weakest where we need it. The whole-response selection ablation tests this directly with our own verifier, and the region-stratified audit shows whether a holistic judge's errors are themselves geographically skewed.

Why not just verify claims with an object detector, as Pelican does? The Dollar Street and GeoDE results imply that detectors will miss everyday low-income and non-Western objects. Standard models reach only 18.0% top-5 accuracy on the lowest income quartile versus 57.4% on the highest, and 37.3% on GeoDE dustbins. A detector-based checker would flag true claims as false in exactly the regions we care about. The verifier audit quantifies this per region rather than assuming it.

Why not just use self-revision at inference time? Volcano more than doubles latency, from 2.7s to 5.8s per query. A model can only catch errors in its own descriptions if it already knows the concept. Including Volcano as a baseline shows whether revision closes any of the regional gap, or whether the gap persists because the knowledge is missing.

Proposed Experiments

The primary metric is claim-level precision on held-out GeoDE and Dollar Street homes, reported per region and income quartile. The headline number is the gap between the best and worst cell.

The decisive checks are four. First, atom-targeted selection must shrink that gap more than random-balanced at matched budget. Second, the improvement must be concentrated in the targeted claim types rather than spread uniformly; a uniform lift would suggest a generic caption-quality effect rather than targeting. Third, the ethnicity-swap accuracy drop on MIXCUBE, which averages 7.64% in the original paper, must fall without training on any MIXCUBE images. GeoDE contains no identifiable people, so this is a pure transfer test, and a null result there would be informative in its own right. Fourth, POPE, FGHE, and MMHal-Bench must show no regression, confirming that geographic gains do not cost general grounding.

We also report the verifier audit and the detector-versus-verifier comparison as standalone findings, since they determine how far any claim-level geographic measurement can be trusted.