Many teams now grade AI outputs with another AI. A multimodal large language model (a chatbot that reads images as well as text) looks at a picture, reads a candidate model's description of it, and assigns a score. This proposal asks a narrow question. Suppose the description gets something wrong about a person in the picture, such as calling a woman in an operating room a "nurse" when the image shows her leading the surgery. Is the judge more willing to let that error pass when it fits a social stereotype? And is it more willing when the person's gender reaches the judge through pixels than when the same fact is stated in words?
If both answers are yes, the judges used to rank, reward, and filter vision-language models are quietly teaching those models that stereotyped mistakes are cheap. The bias would then be amplified at the evaluation step, which no fairness benchmark currently audits. The study builds matched "twin" test items in which the descriptive error is held fixed and only the depicted person's demographics change. It measures how much more often judges excuse stereotype-congruent errors, compares delivering the demographic cue as an image versus as a caption, and checks whether judges ever admit the cue mattered. It then tests whether a small amount of twin-consistency training removes the gap more effectively than the cheaper fixes a reviewer would propose.
Core Claim
Our hypothesis is that multimodal judges show stereotype-congruent error tolerance. Holding the text of an erroneous response constant, the probability that the judge accepts or under-penalizes the error is higher when the image depicts a person for whom the error matches a stereotype. Concretely, define
$$\Delta_m = P(\text{error accepted} \mid \text{congruent}, m) - P(\text{error accepted} \mid \text{incongruent}, m),$$
where $m$ is the channel carrying the demographic cue: the image itself, or a text caption that states the same demographics. We define cross-modal amplification as
$$A = \Delta_{\text{image}} - \Delta_{\text{text}}.$$
The claim has three parts. First, $\Delta_{\text{image}} > 0$. Second, $A > 0$, meaning the bias is larger when it enters through vision. Third, the judge's written justifications almost never mention demographics, even when the cue flips the verdict.
The prediction that $A > 0$ is motivated by work on vision-language fusion. Images Speak Louder than Words used causal mediation analysis on the GLIP object detector, a method that measures how much of an effect flows through each internal component. It found that image features carried 32.57% of gender bias on MSCOCO versus 15.48% for text features, and that deep fusion layers mediated about 57% of encoder bias. That study covered object detection, not judging.
Where This Sits
Three lines of judge research motivate the question, and none of them covers demographics.
MLLM-as-a-Judge benchmarks multimodal judges and documents several failure modes:
- High-score clustering: judges bunch their scores at the top of the scale.
- Verbosity bias: judges reward longer answers.
- A CoT paradox: step-by-step reasoning makes some judges invent visual details and then rely on them.
- A text-description advantage: judges are sometimes more consistent when given a text description instead of the image.
It never varies who appears in the image.
Perception-Judge identifies "response anchoring." A judge that can read a chart correctly on its own still accepts a fluent answer that misreads it. The authors fix this with reinforcement learning on ranked correct, perceptually corrupted, and fully corrupted answers, but their corruptions are limited to counts, colors, and chart values. We borrow their perturbation-and-ranking method and point it at social content. Anchoring predicts that judges should be most easily misled when the false claim is also the statistically expected one.
The Silent Judge shows that text judges change verdicts in response to provenance and recency labels. It measures this with two quantities:
- Verdict shift rate: how often the cue flips the verdict. GPT-4o reached +30% for recency on ELI5.
- Cue acknowledgment rate: how often the written justification mentions the cue. This was 0% across all experiments.
We reuse both metrics. The difference is that our cue is a visual demographic attribute that arrives implicitly, not an explicit text label.
On the bias side, IRIS shows that unified multimodal models answer questions fairly but stereotype heavily when generating images. It traces the amplification to the layers that project image features into the language model. Sathe and Jain show that open vision-language models often score below 0.3 on a neutrality metric, while GPT-4V and Gemini Pro Vision score above 0.9. Neither paper asks how a model behaves when grading another model's output.
These results suggest a constraint on our study. Strong proprietary models may look neutral on direct questions and still show bias as judges. A null result on direct probing therefore says nothing about judging, which is why judging must be tested directly.
Research Program
The test set, which we call Twin-Judge, is built from IRIS occupation images (IRIS-Ideal-52, 52 occupations drawn from labor statistics). Demographic tags come from IRIS's ARES classifier, which reports 88% accuracy, and are confirmed by human spot-checks.
Each twin unit has two images of the same true occupation O whose subjects differ in perceived gender, or in skin tone or age for the secondary axes. Both images are paired with the same erroneous response, which misdescribes the subject's role as O′. We choose O′ so that the error is stereotype-congruent for one twin and incongruent for the other. For example, two surgeon images, one of a woman and one of a man, are each described as showing "a nurse assisting." Each unit also includes a correct response and a response with a non-social perceptual error, such as a wrong object count, as an internal calibration check.
The text of the error is identical across twins, so any difference in how the judge treats it must come from the depicted person. The humanoid-robot images from Sathe and Jain show the same occupational scenes with no human demographics. They serve as a control in which $\Delta$ should be zero.
Each judge is evaluated in three formats taken from MLLM-as-a-Judge: absolute 1–5 scoring, pairwise comparison of correct versus erroneous responses with order swapping, and batch ranking. Each judge also sees each item through three cue channels:
- Image: the raw picture.
- Oracle caption: a human-written description that names the demographics and the scene evidence.
- Captioner route: an automatic description produced by an open captioning model.
The difference between image and oracle caption estimates $A$. The captioner route tests the practical "describe first, then judge in text" pipeline. For every flipped verdict we compute the cue acknowledgment rate from the justification.
On one open judge (Qwen2.5-VL-7B), we run a reduced causal mediation analysis in the style of Weng et al.. We patch the image tokens covering the person's face and body between twins and separately patch the output of the projection layer. This asks whether the judge's demographic sensitivity is localized where IRIS found generation bias. The analysis explains the mechanism but is not the main deliverable.
The proposed fix, Twin-Consistency training, starts from the GRPO recipe of Perception-Judge. GRPO is a reinforcement learning method that scores a group of sampled outputs against each other. We keep their ranking reward (correct above perceptual error above full corruption) and add one term. That term penalizes the absolute difference between the scores the judge assigns to the two halves of a twin pair on the same response. Training uses occupations disjoint from the test occupations, so the judge cannot simply memorize which occupation pairs are tricky.
Baselines
Unmodified judges. The reference point is the judge with no intervention. We run GPT-4o, Gemini, Qwen2.5-VL-7B, and LLaVA-1.6-34B with the standard MLLM-as-a-Judge rubric prompt and order swapping. Any fix must reduce $\Delta_{\text{image}}$ relative to these.
Perception-Judge. This is the strongest existing method for the underlying failure, a judge accepting a fluent answer that contradicts the image. Its paper reports 85.7% perceptual-error identification accuracy for Perception-Judge-Flex-7B, versus 69.5% for base Qwen2.5-VL-7B, on PPJD. We reproduce its training on PPJD with the same 7B backbone, the same GRPO steps, and the same sampling budget as our variant.
This baseline answers whether generic perceptual robustness already closes the stereotype gap. The comparison isolates our key ingredient, because the only difference between the two training runs is whether demographic twins and the consistency term are included. Beating it means a significantly smaller $\Delta_{\text{image}}$ on held-out occupations, measured with paired bootstrap confidence intervals over twin units. It also means PPJD accuracy within about one point of 85.7%.
MR. Judge. MR. Judge trains judges with multiple-choice comparative reasoning and reports surpassing GPT-4o by 9.9% on VL-RewardBench. We run it as released, because reasoning-heavy judges might either reduce or increase stereotype tolerance.
Debiased prompting. We add an explicit fairness rubric and required chain-of-thought to the strongest unmodified judge. No reported result found on Lacuna for demographic error-tolerance gaps under any of these methods. All $\Delta$ values will therefore be new measurements.
Ablations. Two ablations separate the contribution of each training ingredient. One uses twins without the consistency term. The other uses the consistency term on non-demographic perturbations only.
Why not just…?
Why not just give a text-only judge an image description, as the text-description advantage suggests? This is a real threat. MLLM-as-a-Judge found that descriptions sometimes improve judging consistency. The oracle-caption channel tests this directly. If $\Delta_{\text{text}}$ is near zero and the captioner route matches it, the honest conclusion is that practitioners should route through text, and Twin-Consistency training is unnecessary.
There are two reasons to expect a weaker outcome:
- An automatic captioner may itself write "a nurse," which bakes the stereotype into the description. Sathe and Jain's low neutrality scores for open models make this plausible, and the captioner route measures it.
- Text judges still follow silent cues, as The Silent Judge shows.
Why not just blur faces or gendered features before judging? Weng et al. cut detector bias by 22.03% on MSCOCO by blurring gender features. In our task, the evidence for role (scrubs, position at the table) is mostly separate from gender, so blurring could genuinely work. We therefore include a blurred-input condition. The risk is the out-of-distribution artifacts that Sathe and Jain report, along with lost evidence. We will measure accuracy on the non-social calibration items as well as $\Delta$. If blurring matches our training on both, we will report it as the cheaper fix.
Why not just tell the judge to be fair? The fairness-rubric baseline answers this. Two prior findings make a full fix unlikely. The 0% cue acknowledgment in The Silent Judge suggests the judge would not notice its own reliance on the cue. The CoT paradox suggests step-by-step reasoning can add invented visual premises. A drop in $\Delta$ with no loss of accuracy would nonetheless count as a success for prompting.
Decisive Checks
The study rests on six checks:
- Existence: $\Delta_{\text{image}}$ must be significantly positive for human subjects and indistinguishable from zero on the robot control. Otherwise the effect is not about demographics.
- Amplification: $A > 0$ must hold for most judges, with its sign stable across scoring, pairwise, and ranking formats.
- Silence: the cue acknowledgment rate on flipped verdicts should be near 0%. That would extend The Silent Judge's finding to implicit visual cues.
- Mechanism: patching the projection layer should account for more of the twin difference than patching the person-region image tokens. This would echo IRIS's localization and Weng et al.'s fusion result.
- Fix: Twin-Consistency training must beat reproduced Perception-Judge on held-out-occupation $\Delta$.
- No collateral damage: the trained judge must hold PPJD accuracy, VL-RewardBench accuracy, and roughly 79% pairwise human agreement on MLLM-as-a-Judge, the figure reported there for GPT-4V. We use the HQ subset for the agreement check.
If check 1 fails, the paper becomes a useful negative result showing that judges are more robust than generators, a contrast with IRIS's generation gap. If check 5 fails against blurring or captioning, the cheaper method becomes the recommendation.
References
- Images Speak Louder than Words: Understanding and Mitigating Bias in Vision-Language Model from a Causal Mediation Perspective
- MLLM-as-a-Judge: Assessing Multimodal LLM-as-a-Judge with Vision-Language Benchmark
- Mitigating Perceptual Judgment Bias in Multimodal LLM-as-a-Judge for Mathematical and Scientific Reasoning Evaluation
- The Silent Judge: Unacknowledged Shortcut Bias in LLM-as-a-Judge
- Fair in Mind, Fair in Action? A Synchronous Benchmark for Understanding and Generation in UMLLMs
- A Unified Framework and Dataset for Assessing Societal Bias in Vision-Language Models
- MR. Judge: Multimodal Reasoner as a Judge