Crop-Verified Judges for Hallucination Detection in Visual Document RAG

When a vision-language model answers a question from retrieved document pages (screenshots of reports, papers, or webpages with charts and tables), someone has to check whether the answer is actually supported by those pages. Increasingly that checker is another model, an "LLM-as-a-judge" that reads the evidence and issues a verdict. Multimodal judges have a specific failure: when asked to reason step by step, they invent visual details, such as a bar height or a table cell that does not exist, and then build their verdict on the invented detail. The MLLM-as-a-Judge benchmark (MLLM-as-a-Judge) calls this the "CoT paradox": chain-of-thought reasoning makes these judges worse. We propose a judge whose reasoning steps must each point to a box on a retrieved page. Every pointed-to region is then cropped and re-read in a fresh context, separate from the judge's own reasoning. The verdict is computed only from the steps that survive this check. The judge cannot rely on a detail it made up, because a made-up detail fails the crop check and is discarded before the verdict is formed.

In two sentences: a hallucination judge for visual documents should not be trusted to read the page while it is also arguing about it. We therefore make it cite a region for each claim, check each region independently at full resolution, and derive the label mechanically from the checked evidence. This separates reading from judging. We expect the separation to restore the benefit of step-by-step reasoning for multimodal judges and to make their explanations faithful by construction.

Core Claim

The closest existing work falls into three strands, and none of them does this. Halu-J (Halu-J) trains a 7B text judge to sort, reorder, and analyze retrieved passages before labeling a claim, reaching 91% accuracy on ME-FEVER against GPT-4o's 83%. It is text-only, and its critique is never checked against the source. VISA (VISA) and Look as You Think (Look as You Think) teach answer generators to emit bounding boxes on screenshots, at the final answer and at every reasoning step respectively. Both are generators, not judges, and in both the box is an output to be scored rather than a gate the reasoning must pass through. MR. Judge (MR. Judge) shows that reasoning-trained multimodal judges can beat GPT-4o, but it works on single images with no retrieval and no evidence localization.

Our claim is specific and falsifiable. Take a judge that emits box-grounded steps. If its verdict is computed only from steps that an independent crop-level reader confirms, it will (a) outperform the same judge whose verdict is sampled freely after the same grounded reasoning, and (b) show a much lower rate of hallucinated visual details inside its reasoning. The crop check is the key ingredient. Box-grounded reasoning alone, which is what LAT already provides, is the ablation we must beat.

Research Program

The judge receives a question, a candidate answer produced by some other RAG system, and the top retrieved page screenshots. It first splits the candidate into atomic claims, following the subclaim decomposition of MiRAGE (Seeing Through the MiRAGE), which found that claim-level factuality precision tracks human quality judgments while text-centric metrics do not. For each claim, the judge emits a page index, a box, a short literal transcription of what it believes the box shows (for example "the 2019 bar reads 42"), and a relation: supports, contradicts, or no evidence.

The verification step crops each box at native resolution and passes only the crop and the transcription to a reader. The reader is the same backbone, run in a fresh context with no access to the candidate answer or the judge's reasoning. It answers one question: does this crop show what the transcription says? A step that fails is discarded, and the judge may re-search for that claim once. The final label follows a fixed rule: the answer is hallucinated if any claim is contradicted by a verified step, or if a claim has no verified supporting step after the retry.

Because the label is a deterministic function of checked evidence, the explanation cannot diverge from the decision. That is exactly the divergence The Silent Judge (The Silent Judge) documents, where judges shift verdicts on superficial cues while acknowledging those cues 0% of the time. The fresh-context reader also cuts the error compounding that follows once a hallucinated token enters a reasoning chain (Holt et al.). Cropping addresses a practical limitation as well: the judge can reason over downsampled pages while verification reads small chart text at full resolution. VISA identifies high-resolution multi-page context as a memory bottleneck.

Training data comes from the VISA suites (Wiki-VISA, Paper-VISA, FineWeb-VISA), which provide gold evidence boxes. We create hallucinated candidates by perturbing gold answers in the ways the hallucination literature describes: swapping entities, changing numbers read from charts and tables, inverting relations, overstating hedged claims, and adding fabricated but fluent claims. This follows the flawed-candidate synthesis idea of MR. Judge. The gold boxes supervise the step targets. We first fine-tune Qwen2.5-VL-7B on these traces. We then apply a GRPO stage, as in LAT, whose reward combines verdict correctness with the fraction of steps that pass the crop check and overlap a gold box.

The gold-overlap term matters. Without it, the judge could game verification by transcribing vague, trivially true content. We also require transcriptions to contain the specific value or entity named in the claim.

Baselines

The naive reference is direct labeling. Qwen2.5-VL-7B and GPT-4o are shown the pages and the candidate and output CORRECT or HALLUCINATED with no rationale. The MLLM-as-a-Judge benchmark suggests this suffers high-score clustering. It still sets the floor any reasoning method must clear, and it is the cheapest option.

The strongest existing text-side competitor is Halu-J run on the same pages after OCR and VLM captioning of figures and tables, the "semantic translator" route used in MMDocIR-style indexing (MMDocIR). Halu-J's reported result is 91% accuracy and 68.11% evidence-matching rate on ME-FEVER (Halu-J). It gets the same retrieved pages and the same captioner budget as our method.

The strongest multimodal judge competitor is MR. Judge-style reasoning, retrained on our synthetic data so that data is not the difference. Its reported result is surpassing GPT-4o by 9.9% on VL-RewardBench (MR. Judge). We report no result for it on our document-judging task because none exists on Lacuna.

The strongest grounding competitor is LAT's box-grounded chain-of-evidence repurposed as a judge. It reports 8.23% soft exact match and 47.0% [email protected] average gains over vanilla Qwen2.5-VL-7B on Paper- and Wiki-VISA (Look as You Think). We use it in two ways: running a LAT-trained model to regenerate an answer that is compared with the candidate, and training our judge format without the crop gate.

The decisive ablation is the last one. It uses the same backbone, data, and RL steps, with identical grounded steps, but the verdict is sampled freely instead of computed from verified steps. All trained systems share the backbone, the synthetic data, and a matched GPU-hour budget, with tuning sweeps of equal size.

Beating the field means three things. On the perturbed-VISA judging set, our judge must be at least as accurate as the captioned Halu-J pipeline overall and clearly better on chart and table claims. It must beat the ungated ablation by a margin larger than seed variance. It must also cut the rate of hallucinated details in reasoning, measured against gold boxes and human audit, by at least half relative to the ungated ablation.

Why not just…?

Why not just caption the pages and run a text hallucination judge? This is the most serious threat. The MLLM-as-a-Judge benchmark reports that multimodal judges do better with a detailed text description than with raw pixels. VISA, however, found that text pipelines break when OCR misreads technical terms, and MiRAGE found text-centric metrics correlate negatively with human judgments on multimodal evidence. Neither shows that result for a strong captioner feeding a trained text judge. We therefore stratify results by evidence type. If the captioned pipeline matches us on charts and tables, the pixel route is not worth its cost, and we would report that.

Why not just use a LAT-trained model, since it already grounds every step in a box? LAT's boxes are scored during training but never gate its conclusions at inference. A box can be correct while the transcription of its contents is wrong, and that is precisely the CoT-paradox failure. The ungated ablation isolates this. If gating adds nothing over grounded steps alone, our central claim fails.

Why not just prompt a frontier model directly for a label? Direct labeling is fast. The evidence we have points against it: GPT-4o trails a trained 7B judge on ME-FEVER (83% vs 91%), zero-shot Qwen2-VL-72B locates evidence boxes only about 1.5% of the time on Wiki-VISA, and frontier judges shift verdicts on unacknowledged cues. We still run it, because on easy text-heavy pages it may be competitive, and we report the accuracy-per-latency tradeoff honestly.

Proposed Experiments

The primary benchmark is perturbed-VISA judging. We take held-out Wiki-VISA, Paper-VISA, and FineWeb-VISA questions and pair them with gold answers, perturbed answers, and real outputs from VISA- and LAT-style generators, with human labels on a subset of about 1,500 real outputs. Metrics are verdict accuracy and F1, box IoU of cited evidence, and step-hallucination rate, meaning steps whose transcription disagrees with the gold region. We report results in both single-page and multi-candidate settings. The multi-candidate setting matters because VISA's own box accuracy falls from 54.2% to 32.3% when multiple candidates are retrieved.

Three decisive checks follow.

The first tests the CoT paradox directly. On the MLLM-as-a-Judge pair-comparison split, we treat the single image as the retrieved page and compare direct judging, free CoT, and crop-verified CoT with one backbone. GPT-4V's reported human agreement there is about 79%. The prediction is that free CoT underperforms direct judging, reproducing the paradox, while crop-verified CoT exceeds both.

The second is a modality-cost check. We render ME-FEVER's text evidence as screenshots and compare our judge with Halu-J's reported 91%, which measures what the visual route costs when layout carries no information.

The third is a visual shortcut test in the spirit of the Silent Judge. We hold content fixed while moving the evidence page later in the retrieved list, downsampling it, or adding a professionally styled distractor page. We measure verdict shift rate and cue acknowledgment. The prediction is that ungated judges shift silently, following VISA's first-page bias, while the gated judge shifts only when the verified evidence changes.