acceptodds
Under review as a conference paper at ICLR 2027

OmniFair: A Unified Fairness Benchmark Across Tasks and Modalities

Abstract

Fairness evaluation of multimodal large language models (MLLMs) remains fragmented across tasks and modalities. Existing instance-aligned multi-task benchmarks remain text-only, whereas instance-aligned multimodal benchmarks remain confined to individual tasks, preventing controlled separation of task and modality effects. We introduce the Bias Semantic Unit (BSU), a five-dimensional representation that separates stereotype semantics from task- and modality-specific realizations. From 143 fairness benchmarks and 11M multilingual news articles, we construct BiasAtlas with 416K BSUs. Projecting 3,075 representative BSUs across five task paradigms and three modalities yields OmniFair, a benchmark of 46,125 instances with joint cross-task and cross-modal semantic-source alignment; human validation finds 97.9% semantic fidelity and 95.5% multimodal media effectiveness. Across 15 state-of-the-art MLLMs, instance-level discrimination decisions show almost no agreement across tasks () and limited agreement across modalities (). Category-level bias-rate rankings are likewise nearly uncorrelated across tasks (mean pairwise across semantic axes) but more consistent across modalities (). This contrast shows that task paradigms reshape the semantic risk landscape, whereas modalities alter which individual stereotypes elicit discrimination while preserving more of the category-level structure.

Primary area:
datasets and benchmarks
Keywords:
fairness, bias evaluation, multimodal large language models, benchmark, cross-modal, cross-task, stereotype

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_AgfoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors' primary research focuses on how the same social stereotype manifests in fairness evaluations across different tasks and modalities; the paper is essentially a benchmarking study. Specifically, the authors first clarify that the research scope focuses on "stereotype semantics." To enable cross-task and cross-modality evaluation, they designed the Bias Semantic Unit (BSU), which encompasses contexts, protected attributes, target groups, bias expressions, and harm categories. Crucially, the BSU decouples stereotype semantics from specific evaluation tasks and input modalities, allowing a single BSU to generate corresponding evaluation instances across various settings. After organizing these units, the authors constructed an evaluation dataset derived from two sources: real-world news and existing academic fairness benchmarks. They built "BiasAtlas" through LLM-assisted extraction, normalization, and deduplication, selecting 3,075 representative BSUs to generate the "OmniFair Benchmark." For each BSU, the authors generated test cases covering five task paradigms and three input modalities (resulting in 15 evaluation instances per BSU), yielding a total of 46,125 evaluation instances (3,075 × 15). Subsequently, 15 models were tested on their supported tasks and modalities, with results converted into binary classifications (biased vs. unbiased) based on task-specific rules; this process yielded the paper's core evaluation results. Additionally, the authors validated the benchmark's construction quality and the reliability of the evaluation results through human annotation and consistency checks involving LLM judges.

Based on the experimental results, the paper presents 12 findings across six key areas: differences in bias rates across models, tasks, and modalities; changes in bias when a unified BSU is applied to different tasks; near-zero cross-task correlations but higher cross-modal correlations in category-level bias-rate rankings (context, protected attribute, and harm category); the relationship between bias and the direction of the stereotype; differences in bias between real-world news and academic datasets; and differences in observed bias between explicitly spoken group identities and identities conveyed solely through voice timbre.

Critical strengths
  1. Cross-task and cross-modal semantic-source alignment: The proposed BSU (Bias Semantic Unit) and its associated cross-task, cross-modality projection scheme offer both novelty and practical value. The authors attempt to abstract stereotype-related semantics from fairness evaluations into a unified representation—independent of specific tasks or modalities—enabling the generation of corresponding test instances for the same stereotype across different evaluation settings. This design establishes a unified foundation for tracking model performance regarding a specific bias semantic under varying evaluation conditions and offers significant potential for scalability.

  2. Comprehensive data construction and quality validation: The paper involves extensive data engineering, encompassing 143 academic benchmarks, approximately 11.4 million news articles, 416K unique BSUs, and 46,125 final test instances, thereby providing substantial data support to the fairness research community. Furthermore, for components involving LLM assistance, the authors employed human annotation, verified "Judge" consistency, and conducted robustness analyses on specific metrics, providing supporting evidence for the quality of the constructed benchmark.

Critical weaknesses
  1. Insufficient justification regarding the comparability and scientific significance of cross-task fairness evaluation results: One of the paper's core findings is that the same BSU yields inconsistent bias assessments across different tasks and modalities. However, while the authors align the semantic source of the test content (the same BSU), the specific fairness-related behaviors being measured remain task-specific. For instance, consider a BSU concerning the technical competence of older employees: the OG task measures whether the model generates stereotypical content under subtle elicitation prompts; the MCQ task measures whether the model relies on group identity to make judgments when information is insufficient; the DT task measures the model's ability to recognize stereotypes; and the PC task measures whether the model assigns different toxicity, offensiveness, or bias scores to otherwise identical statements referring to different groups. These tasks not only require distinct model behaviors and capabilities but also rely on markedly different operational definitions of bias. Consequently, even though these tasks share the same BSU, inconsistent results do not necessarily indicate inconsistency in the model's underlying fairness performance, nor do they necessarily constitute a novel scientific discovery. I believe a primary issue with the paper is that, while it possesses a solid core proposition and framework (BSU), it attempts to address too many questions; the interpretation of cross-modal and cross-task differences remains insufficiently clear, and the cross-task consistency analysis reduces task-specific outcomes to binary bias labels—a metric far too coarse for the scope of the investigation. The authors need to better justify the rationale behind these cross-task comparisons and articulate what new insights regarding fairness research can be gleaned from the observed inconsistencies.

  2. The depth and novelty of the experimental findings are relatively limited: Although the paper reports 12 findings covering tasks, modalities, model scales, "Thinking" configurations, and data sources, most of these remain descriptive—focusing on differences in bias rates and consistency across conditions. The paper lacks in-depth analysis regarding the underlying causes of these differences or the broader principles of fairness they might reveal. For instance, variations in cross-modal bias rates could stem from a combination of model perception capabilities, refusal behaviors, and actual bias manifestations. Although Finding 12 discusses perception failure and Appendix I.4 analyzes refusal rates, I find their implications for the broader cross-task and cross-modal results insufficiently developed. Consequently, I feel the extensive experimental coverage has not yet translated into sufficiently profound scientific insights regarding AI fairness.

  3. There is room for improvement in the paper's presentation: In terms of structure, the paper reads more like a technical report; the authors devote considerable space to listing the 12 findings individually, yet the connections between them are weak, causing the core narrative and key scientific conclusions to lose prominence. Regarding the figures, the text in some is too small, and certain elements lack clarity even when zoomed in, which somewhat detracts from the reading experience.

Questions for authors
  1. Regarding the comparability of cross-task evaluations: Could the authors further clarify what semantic-source alignment makes comparable across tasks and what conclusions about model fairness can be drawn from the observed differences? Given the differing operational definitions of bias across OG, MCQ, DT, PC, and CT, how do the authors distinguish between inconsistencies arising from differences in the tasks' measurement objectives and genuine, scientifically significant differences in model fairness? This distinction seems crucial for understanding the paper's core conclusions.

  2. Interpretation of key findings: What concrete implications do the observed cross-task and cross-modal differences have for the design and interpretation of future fairness evaluations? In particular, how should practitioners use these findings when choosing evaluation paradigms or interpreting aggregated bias rates?

Reviewer_CoqsOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces OmniFair, a unified benchmark for evaluating fairness in multimodal large language models across different tasks and modalities. By introducing Bias Semantic Units (BSUs), the authors enable consistent comparisons across tasks and modalities. Extensive experiments on MLLMs reveal significant inconsistencies in fairness evaluations across tasks and, to a lesser extent, modalities.

Critical strengths

The paper addresses a timely and relevant problem with a well-motivated and clearly structured approach. The proposed benchmark is carefully designed and extensively validated, providing valuable insights into fairness evaluation across tasks and modalities. The experimental evaluation is comprehensive and well organized.

Critical weaknesses

None. I did not identify any significant weaknesses.

Questions for authors

None. The paper is clear, and I have no questions that would materially affect my assessment.

Reviewer_wWgvOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents OmniFair, a large-scale, unified fairness benchmark designed to evaluate stereotype-based bias in multimodal large language models (MLLMs) under joint cross-task and cross-modal semantic-source alignment. Prior fairness benchmarks either evaluate a single task across modalities or evaluate multiple tasks strictly in text, confounding the source stereotype with observation conditions (task framing and input modality). To decouple bias semantics from evaluation formats, the authors introduce the Bias Semantic Unit (BSU), a five-dimensional representation that separates stereotype semantics from task- and modality-specific realizations. Using this abstraction, the authors build BiasAtlas, a repository of 416,690 unique BSUs compiled from two complementary sources: 11.4M multilingual news articles (capturing real-world reported discrimination events) and 143 academic fairness benchmarks (derived from an initial survey of 454 papers). The authors conduct extensive quality validation of the benchmark artifacts alongside an evaluation of 15 frontier models.

Critical strengths
  • A chronic limitation in existing MLLM fairness evaluations is the conflation of the underlying stereotype with the observation format (such as multiple-choice framing, prompt formulation, or input modality). Existing suites either evaluate multiple tasks purely in text or benchmark modalities within isolated, single-task configurations, making it impossible to ascertain whether measured disparities stem from the stereotype itself or the task/media container. The paper addresses this fundamental challenge by proposing the 5-dimensional Bias Semantic Unit (BSU), instantiating 3,075 representative BSUs across a fully crossed 5×35 \times 3 grid (46,125 instances) with shared counterfactual stimuli and underlying media. This controlled semantic-source alignment represents a substantial conceptual and infrastructural advance that enables the first rigorous, controlled separation of task and modality effects in multimodal fairness.
  • The paper delivers vital, counter-intuitive empirical findings that directly challenge the validity of single-task and single-modality evaluations. By evaluating 15 modern LLMs and MLLMs on aligned instances, it demonstrates that instance-level discrimination decisions exhibit virtually zero agreement across tasks (κ=0.02\kappa = 0.02) and category-level risk profiles are reshuffled (ρ≈0\rho \approx 0), whereas modality shifts preserve substantially more category-level risk structure (ρ∈[0.48,0.59]\rho \in [0.48, 0.59]). This insight proves that ubiquitous single-task benchmarks (e.g., BBQ multiple-choice QA) cannot serve as reliable proxies for general model fairness. Moreover, the empirical claims are backed by exemplary methodological rigor, explicitly diagnosing and disentangling acoustic perception failures in timbre, judge-model dependencies, option-order artifacts, and refusal-masking dynamics.
Critical weaknesses

The central empirical conclusion that model discrimination decisions show almost no agreement across tasks (κ=0.02\kappa = 0.02) and category-level risk rankings are uncorrelated (ρ≈0\rho \approx 0), is largely an artifact of comparing fundamentally incommensurate behavioral capabilities and binarization criteria under the umbrella of "bias rate":

  • DT measures a model's analytical capability to detect subtle bias in an explicit statement (where bias is defined as a detection miss rate on positive instances).
  • OG and CT measure generative compliance versus guardrail resistance when subjected to subtle adversarial red-teaming or persona prompts designed to bypass safety filters (scored ≥3\ge 3 by an LLM judge).
  • MCQ measures syllogistic reasoning under ambiguity in a BBQ-style forced choice ("Cannot be determined" vs. stereotypical options).
  • PC measures independent scoring calibration stability, where any non-zero difference (∣Δ∣≥1\vert{}\Delta\vert{} \ge 1) on a 5-point scale between two separate rating prompts is classified as discriminatory.

Because an analytical failure to flag an implicit stereotype (DT) shares little underlying model mechanics with generative degeneration under adversarial prompting (OG) or prompt-rating noise (PC), near-zero Cohen's κ\kappa and near-zero ranking correlations across these tasks are inevitable by construction. Treating these distinct task objectives as interchangeable observation windows of the same latent bias decision overstates the conceptual divergence of model behavior and risks mistaking metric incommensurability for a substantive reshaping of the "semantic risk landscape".

Questions for authors
  • The OG projection relies on an LLM generating one of six subtle elicitation strategies (e.g., persona roleplay, advice seeking, devil's advocate) to bypass guardrails. In contrast, DT, MCQ, and PC use fixed, structured phrasing templates. How much of the divergence in bias rates and category rankings between OG and the rule-based paradigms is driven by variations in the prompt-elicitation strategy chosen for a given BSU? Did the you test whether varying the prompt strategy for the same BSU in OG shifts the bias outcome?
  • The authors observe that the reduction in bias under vocal timbre is largely an artifact of perceptual failure that models fail to extract demographic signals (such as age or gender) from the acoustic timbre and default to treating the speaker as an unspecified individual (Finding 12). A parallel threat applies to visual projections relying on scene context (Strategy 2) or in-image text/placards (Strategy 3): While human annotators verified that the media assets convey identity (95.5% media effectiveness), did you conduct an explicit manipulation check on the evaluated models (e.g., a neutral probe asking the model to identify the demographic attribute depicted in the image or audio)? What proportion of image instances are actually perceptually grounded by open-source and smaller models before responding to the bias evaluation task?
1 comment
AuthorsOpenReview ↗

Reviewer–Author Discussion Period Will End Soon

Dear Reviewer wWgv,

The reviewer–author discussion period will end soon. We would be grateful to hear whether our responses and revised manuscript address your concerns, and we remain happy to clarify any unresolved questions.

Your feedback prompted us to sharpen the paper substantially. We now frame OmniFair as a semantically aligned benchmark for testing whether task-local fairness conclusions serve as empirical proxies across evaluation forms, rather than as interchangeable measurements of a latent fairness capability. We also added the paired Direct OG and neutral grounding controls requested in your questions, incorporated controlled task comparisons and extended consistency analyses, and reorganized the manuscript around this focused claim.

At its core, OmniFair asks whether a favorable fairness result in one task or modality provides evidence about how the same discrimination semantics will manifest in another. We believe this question matters because models serve people across many forms of interaction, while harm does not cease to matter when its technical form changes.

We sincerely appreciate your time and consideration.

Warm regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.