Source-linked AI summary
Melange: A Synthetic Visual Bias Benchmark for Multi-modal Large Language Models
Razieh Rezaei, Ehsaneddin Asgary, Masoud Jalili Sabet
TL;DR
MLLM bias benchmarks have lacked scale, diversity, and visually grounded scenarios that make target groups clear in images. Melange extends BBQ into a seven-category synthetic visual benchmark and evaluates matched text and image settings; switching from text to vision reduces fairness by 0.291 on average.
Problem
Existing MLLM bias resources lack scale, diversity, and comprehensiveness, while visually grounded datasets are needed for meaningful comparison across text and image inputs.
Source: §1 · p. 1Method
Melange extends BBQ into seven visually grounded bias categories, validates generated images with automated VQA and human review, and evaluates models in matched text-only, image-only, and image-with-context settings.
Source: §1 · pp. 1–2, §2.2 · p. 3Results
Switching from text to vision reduces fairness by a mean of 0.291 across 23 open-source MLLMs, with the gap varying by model family rather than scale alone.
Source: §1 · p. 2Takeaways & Limitations
Models recognize depicted demographic attributes at 90–95% accuracy yet still fail to abstain, and biased answers align with human stereotypes roughly three-quarters of the time.
Source: §1 · p. 2All 12,640 scenarios are ambiguous, so Melange measures abstention under ambiguity and cannot detect bias when a correct answer exists.
Source: §7 · p. 9
Abstract
from arXiv · showhide
Multi-modal Large Language Models (MLLMs) have seen broad adoption, but their biases are still far less examined than those of standard LLMs, creating possible social harms when deployed in real-world settings. Prior efforts to extend bias analyses from LLMs into the multi-modal space either cover only a few bias categories or do not ensure that the groups being compared are clearly addressed by visual cues in the images. To overcome these limitations, we introduce Melange, a wide-ranging synthetic visual benchmark covering seven categories for assessing social biases in MLLMs. It is grounded in the widely used language-only BBQ dataset (Bias Benchmark for Question Answering) from BigBench, carefully extending its constructed contexts via controlled generation of visual scenarios so the target attributes appear visually and clearly. To maximize reliability, our generated images pass through two stages of validation: automated screening using visual question answering (VQA) and subsequent manual inspection by human annotators. By systematically evaluating 23 open-source MLLMs across three matched evaluation settings---text-only, image-only, and image-with-context---Melange isolates the fairness effect of the input modality. Switching from text to vision reduces fairness by 0.291 on average, a gap that varies by model family rather than by scale alone. A perception control confirms that this gap is not an artifact of visual comprehension: models identify the depicted attributes at 90--95\% accuracy yet still fail to abstain. When models do commit to a biased answer, polarity-aware stereotype alignment shows they choose the human-stereotypical response roughly three-quarters of the time.
1. Introduction
Melange extends BBQ into a visually grounded benchmark for assessing MLLM bias across seven categories. Its matched evaluation finds a text-to-vision fairness gap that varies by model family, alongside strong visual attribute recognition and frequent human-stereotypical answers.
1. Introduction
- Melange extends BBQ to the visual domain with seven bias categories: age, gender, appearance, socioeconomic status, race, religion, and disability.The benchmark covers 12,640 scenarios and 126,400 evaluation images.Source: §1 · pp. 1–2
- Melange filters BBQ prompts for explicit visual cues, generates images, and applies automated VQA and human validation for prompt-image alignment and semantic consistency.Human validation also removes biases that could propagate from the text-to-image model.Source: §1 · p. 1
- Ten visually diverse but semantically equivalent images are curated for each unique prompt and evaluated alongside text-only BBQ counterparts.This supports direct fairness comparisons across modalities.Source: §1 · p. 2
- Switching from text to vision reduces fairness by a mean of 0.291, with the gap varying by model family rather than scale alone.Across 23 open-source MLLMs, attribute recognition remains 90–95% accurate, while biased answers align with human stereotypes roughly three-quarters of the time.Source: §1 · p. 2
- The benchmark evaluates 23 open-source MLLMs across three matched settings: text-only, image-only, and image-with-context.The matched settings are designed to isolate the fairness effect of input modality, and the evaluation spans nine model families.Source: §1 · p. 2
2. Dataset Generation
Melange builds on BBQ’s social-bias question sets and selects contexts whose target attributes can be clearly represented visually. It creates scenarios by composing separately generated scenes and characters.
2.1. Background: The BBQ Dataset
- BBQ measures social biases across protected classes using ambiguous contexts, where “Unknown” is correct, and disambiguated contexts with a specific correct individual.The dataset contains over 58,000 prompts and covers nine social dimensions plus two intersectional categories.Source: §2.1 · p. 2
2.2. Textual Bias Scenarios: The BBQ dataset
- Melange selects seven BBQ categories with clear visual differentiation, excluding nationality and sexual orientation because these attributes often lack distinct visual cues.Each context appears four times, varying character order and question polarity; contexts without reliable visual cues are excluded.Source: §2.2 · p. 3
- Each context is decomposed into a manually written scene and two character descriptions derived from BBQ answer options, rendered separately and composed into a visual scenario.Figure 1 shows examples from Appearance, SES, and Religion.Source: §2.2 · p. 3, Figure 1 · p. 3
3. Dataset Sanity Checks
Melange constructs scenarios from separately generated visual components, then filters, tags, matches, and human-checks the images. Its VQA pipeline uses token probabilities and a threshold to classify responses.
3. Dataset Sanity Checks
- Scenes, characters, and composite scenarios are generated as separate stages, with the scene manually authored and character prompts derived from BBQ answer options.The image-generation pipeline composes a scene and two characters into each complete scenario.Source: §3 · p. 3
- Each character has 25–50 curated depictions after VQA screening for person count, photorealism, face visibility, and nudity.When fewer than 25 valid images remain, annotators iteratively refine the prompt; Figure 2 summarizes filtering, tagging, human checks, and sampling.Source: §3 · p. 3, Figure 2 · p. 4
- Images are tagged for age, gender, and race, then characters are matched on non-target tags so paired characters differ only on the bias category under study.Annotators also verify realism, intended descriptions, and that tag-based matches introduce no unintended differences; scenarios are sampled from character and scene pools.Source: §3 · pp. 3–4
3.1. Binary VQA via Token Probabilities
- VQA stages classify binary responses from next-token probabilities, treating a yes or no as predicted when its highest-probability first-token variant meets threshold τ.A prediction is correct when the thresholded response matches the binary ground truth; otherwise it is undefined if neither response meets threshold.Source: §3.1 · pp. 4–5
4. Model Evaluations
Across three matched settings, the evaluation finds that fairness drops when models move from text to images, with the size of that drop varying by model family. A perception control shows that models generally identify the depicted attributes despite the fairness gap.
4. Model Evaluations
- The evaluation includes 23 MLLMs from nine families, with parameter counts from 1B to 72B; option-position preference largely cancels because BBQ distributes the unknown option nearly uniformly.The neutral option appears in positions A, B, and C in 32.6%, 34.3%, and 33.1% of cases, respectively.Source: §4 · p. 5
- The three settings progress from text-only BBQ to image-only Melange and image-with-context, which adds a neutral scene description naming neither character.The image settings use spatial answer options; the correct answer is always “unknown.”Source: §4 · p. 5, Figure 3 · p. 6
- FaiR is the fraction of responses selecting a neutral answer, and the text→vision gap and context recovery quantify the fairness change across settings.The gap is MELANGE minus BBQ LLM; recovery is MELANGE WITH CONTEXT minus MELANGE.Source: §4 · p. 5
4.1. Results
- −0.291 mean FaiR gap separates MELANGE from BBQ LLM; every model is less fair in the image setting, with gaps ranging from −0.525 to −0.098.Table 2 reports scores for all 23 models across the three settings.Source: §4.1 · p. 5, Table 2 · p. 7
- The gap varies by family rather than scale: Qwen2.5-VL improves monotonically with size, while the 38B InternVL3.5 has the largest gap at −0.525.In a same-release Qwen3.6 comparison, the dense 27B gap is −0.137, versus −0.414 for the MoE 35B-A3B, which has the best text-only FaiR at 0.956.Source: §4.1 · pp. 5–6
- +0.047 average FaiR recovery follows neutral scene text; models with the largest vision losses recover most, while robust models are flat or slightly negative.The reported largest recoveries include InternVL3.5-14B at +0.157, 1B at +0.145, and 38B at +0.141.Source: §4.1 · p. 6
4.2. Perception Control
- Perception accuracy spans 0.905–0.931 on attribute probes and 0.924–0.953 on detection probes, while FaiR spans 82 points across seven representative models.Restricting FaiR to items the model demonstrably perceives changes it by at most 0.011, slightly downward.Source: §4.2 · p. 6, Table 3 · p. 7
- Religion perception reaches 0.848, although religious affiliation is not visually determinable in principle and images encode visible markers such as clothing or headwear.Thus, religion scenarios partly test responses to visual cues that the benchmark framing treats as absent.Source: §4.2 · p. 6
5. Ablations
The modality gap and category-level bias patterns vary by model family rather than following a universal scale trend. Correcting for question polarity reveals that biased answers usually align with human stereotypes.
Scale effects
- Scale effects are family-dependent: Qwen2.5-VL’s text-to-vision gap shrinks monotonically from −0.345 (3B) to −0.098 (72B), while InternVL3.5-38B has the largest measured gap (−0.525).Among Qwen3-VL models, the 32B is more robust than the 30B-A3B (−0.121 vs. −0.287).Source: §5 · p. 7
Category-level patterns
- Race and religion have the highest FaiR scores, while age and appearance tend to have the lowest; this ordering is stable within families across scales.The reported category pattern is consistent with family-specific visual encoders and alignment procedures shaping bias profiles.Source: §5 · p. 7, Figure 5 · p. 8
Stereotype alignment
- 0.770 average polarity-aware stereotype alignment across 23 models and three settings, compared with a polarity-blind 0.496 that is near 0.50 by construction.Alignment averages 0.90 for negatively framed questions and 0.74 for positively framed questions; it is higher in BBQ LLM (0.853) than MELANGE (0.725) or MELANGE WITH CONTEXT (0.733).Source: §5 · pp. 7–8
Stereotype alignment scope
- Stereotype-alignment analysis covers six categories, excluding SES because group membership cannot be reliably inferred from appearance.Gender annotations cover 96% of scenarios, while race annotations cover 27%.Source: §5 · p. 8
6. Related Work
Prior bias benchmarks primarily assess language-only models, while visual-bias research often focuses on generated imagery or dataset characteristics. Melange instead evaluates MLLM responses to visually presented ambiguous situations using paired text and visual scenarios.
Bias benchmarks for language models
- BBQ [Parrish et al., 2022] and other cited benchmarks assess LLM bias in text, so they cannot establish whether model behaviour changes when social information is conveyed through images.BBQ is Melange’s foundation, preserving ambiguous scenarios and seven bias categories while adapting evaluation to visual inputs.Source: §6 · p. 8
Bias in vision and text-to-image generation
- Prior visual-generation and dataset studies examine demographic bias in generated imagery or dataset construction, rather than downstream MLLM responses to ambiguous visual situations.The passage cites Bianchi et al. (2023), Cheong et al. (2024), and Paullada et al. (2021).Source: §6 · pp. 8–9
Evaluating bias in multimodal models
- Existing MLLM bias evaluations have relatively constrained coverage, and comparable BBQ-to-vision studies may leave paired demographics unmatched or introduce written-name linguistic bias.Melange addresses these issues with spatial references and pairs aligned across demographics except for the target attribute.Source: §6 · p. 9
Synthetic and counterfactual evaluation data
- Melange’s paired text-only and vision-only evaluations of 12,640 BBQ-derived scenarios distinguish it from cited synthetic-data work that lacks corresponding text and vision versions.A perception control is included to show that the observed modality difference cannot be attributed to limited visual competence.Source: §6 · p. 9
7. Limitations
The benchmark and its analyses have boundaries in scenario ambiguity, stereotype-annotation coverage, image realism, reproducibility, and control of visual confounds.
Scenario ambiguity
- All 12,640 scenarios are ambiguous, so Melange measures abstention under ambiguity and cannot assess bias when a correct answer is available.Disambiguated contexts would test whether models apply stereotypes when sufficient information supports a correct answer.Source: §7 · p. 9
Annotation coverage
- Stereotype annotations are absent where group membership is not reliably inferable from appearance: SES coverage is 0%, race 27%, and gender 96%.The stereotype-alignment analysis therefore covers six categories and excludes SES.Source: §7 · p. 9
Synthetic images
- All images are synthetic, and the evaluation cannot isolate how distributional differences from natural photographs affect model behaviour.The perception probe finds a religion score of 0.848 despite religious affiliation being visually indeterminate in principle, suggesting the generator encodes visible religious markers.Source: §7 · pp. 9–10
Reproducibility
- Two original side-by-side character-crop settings are not reproducible from the public release because per-character renders are unavailable.The paper reports results for the three reproducible settings: BBQ LLM, MELANGE, and MELANGE WITH CONTEXT.Source: §7 · p. 10
Visual confounds
- The methodology does not explicitly control latent visual confounds such as facial expressions or clothing colour, which could create bias pathways beyond the intended variable.The passage notes prior work associating bright clothing with positive attributes and identifies control of these factors as future work.Source: §7 · p. 10
8. Conclusion
Melange pairs BBQ-derived scenarios with synthetic images to assess MLLM fairness across modalities. Across 23 open-source models, vision reduces fairness by 0.291 on average, with the gap varying by family rather than scale alone.
8. Conclusion
- Across 23 open-source MLLMs, switching from text to vision reduces fairness by 0.291 on average, with gaps varying by family rather than scale alone.Models identify depicted demographic attributes at 90–95% accuracy yet still fail to abstain; when they commit to biased answers, polarity-aware alignment shows they choose the human-stereotypical response roughly three-quarters of the time.Source: §8 · p. 10