Why multimodal AI judges can be swayed by explanations

A preprint on arXiv reports that a plausible explanation can be enough to mislead a multimodal AI judge when the task requires visual verification. It was reported independently by arXiv cs.CL and arXiv cs.AI. The paper says multimodal large language models increasingly assess answers alongside explanations in training and evaluation pipelines, and introduces a method called JUPE.

This is an unreviewed preprint, and the supplied material does not give its datasets, models, measurements or results. Still, it names a problem worth separating from the particulars of any benchmark: a system asked to check an image can instead be persuaded by a coherent account of the image. A convincing explanation is not the same thing as visual evidence.

A judge has to assess evidence, not only language

A multimodal judge is a language model that receives more than language. In the setting described by the preprint, it assesses an answer and its explanation where visual verification is required. Its task is therefore different from simply generating a fluent image description. It must decide whether the visual material supports the answer being offered.

That assessment contains distinct inputs. There is the answer itself: a claim about what the visual material shows. There is the explanation: the proposed reasoning or account that makes the answer seem justified. And there is the visual evidence. These inputs can agree, but they do not have the same job. The image is what a visually grounded claim must ultimately answer to.

Consider a hypothetical answer that makes a claim about the location or identity of something in an image. Its accompanying explanation may tell a tidy, familiar story about why the claim follows. The story can be internally consistent even when the image does not support the claim. If the judge accepts the candidate because the account is easy to endorse in language, it has substituted a test of plausibility for a test of evidence.

That is the concern in the new paper's title. A preprint on arXiv reports that ordinary plausible explanations may mislead multimodal judges in cases that demand visual verification. The narrowness matters. The paper is not, from the supplied material, a general claim that explanations are always harmful or that multimodal models cannot process images. It is about whether a particular kind of evidence check can be displaced by persuasive text.

Plausibility and verification answer different questions

Plausibility asks whether an explanation hangs together. Does it sound like an account that could fit the answer? Visual verification asks something stricter: does the visual input support the answer? A response can pass the former while failing the latter.

The preprint's abstract explicitly starts from the fact that AI judges are sensitive to textual factors. That sensitivity is not automatically a defect. Language is part of many tasks, and explanations can contain relevant information. The problem arises when textual coherence gets treated as a substitute for checking the evidence the task says matters.

A judge could be affected by a well-phrased rationale even if that rationale adds no visual support. Equally, a sparse but correct answer could be disadvantaged because it gives the judge less language to agree with. Those are different routes to the same failure: the final score looks like a verdict on the image, while the decision may actually reflect the writing around it.

The supplied material does not establish how any model produces that effect. It does not describe an architecture, an attention pattern, or an internal mechanism. It says the authors introduce JUPE to investigate whether plausible explanations sway judgments requiring visual verification. That is the appropriate level of certainty: the paper poses and studies the vulnerability; the available summary does not explain every cause of it.

The useful test holds the image steady

The clean way to isolate this problem is controlled disagreement. Hold the visual material and proposed answer steady, then vary only the explanation. If the explanation has no evidentiary role, a visually grounded judgement should not change merely because the wording becomes more persuasive. The inverse test also matters: hold the prose steady and alter the visual evidence. A judge that is performing visual verification should react to the evidence change.

That sounds simple, but it makes the evaluation question much sharper. It distinguishes a model that can repeat a plausible rationale from a model that is using the image to decide whether the rationale deserves credit. It also reveals why aggregate scores can conceal the issue. A judge may reach an acceptable answer for the wrong reason, and a final score alone does not expose the path that produced it.

JUPE may provide a specific protocol for this kind of test, but the supplied material does not say how it constructs examples, what it varies, or how it measures a misleading judgement. That missing detail limits what can be concluded from the announcement. A reader cannot infer how broad the effect is, whether it transfers across tasks, or whether it applies to a particular model family.

It does establish the question that a good report should answer. What was held constant? What changed? Did the judge respond to visual evidence, explanatory wording, or both? Without those conditions, a claim that a multimodal judge is accurate may be too coarse to be useful.

Explanations cannot simply be treated as noise

The remedy is not necessarily to remove explanations. Whether explanation should influence a decision depends on the task. If a task genuinely asks a judge to assess reasoning as well as an answer, the explanation is part of the object being evaluated. If the task is visual verification, however, the evaluator must be clear about which parts of the explanation are themselves grounded in the image.

That boundary is where systems break. An explanation may restate what is visible, make an unsupported leap, or add information that the image cannot verify. Treating all of those alike makes the score hard to interpret. Conversely, treating every explanation as irrelevant can obscure whether an answer actually follows from the available evidence. The task definition, not a generic preference for longer or shorter rationales, has to decide the role of prose.

A separate preprint on arXiv reports that evaluating LLMs and agentic systems requires combining benchmark accuracy, judge-based assessment, runtime performance, tool-use behaviour and retrieval quality. It also says differences in interfaces, metric meanings and trace formats can hide failures in agentic workflows. That paper does not validate JUPE, but it reinforces the practical implication: a judge score is one signal in an evaluation system, not a complete account of what occurred.

For visual judging, traces should make it possible to tell whether the score changed with the evidence or with the presentation. The material supplied here does not prescribe a reporting format. It does make clear why the format matters: a pipeline cannot diagnose a text-induced error if it records only a final pass or fail.

The preprint narrows what to ask of a judge

The immediate lesson is not that multimodal judges are unusable. A preprint on arXiv reports a more precise concern: plausible explanations can mislead them when they are meant to perform visual verification. The paper's abstract does not identify a universal failure rate, name a best mitigation, or show that every evaluation pipeline has this problem.

That restraint changes the questions a technical reader should ask. What visual claim is the judge supposed to verify? Does the result change when explanatory wording changes but the image does not? Does the evaluation separate an answer's visual correctness from the fluency of the explanation attached to it? Can a failure be traced to evidence, prose, or the task's own ambiguous specification?

Those questions matter wherever a model evaluates another model, whether the score is used during training or reported as an evaluation result. A polished rationale can be useful for humans reviewing a decision. It should not receive credit for seeing. The durable standard is straightforward: when a task requires visual verification, the evaluator must be shown to respond to the visual evidence rather than merely accepting a convincing story about it.