Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FairJudge: MLLM Judging for Social Attributes and Prompt Image Alignment

Zahraa Al Sahili, Maryam Fetanat, Maimuna Nowaz, Ioannis Patras, Matthew Purver

TL;DR

FairJudge introduces a calibrated evaluation framework that treats instruction-following multimodal LLMs as fair judges for text-to-image (T2I) assessment. By constraining outputs to a closed label set, requiring explicit visual evidence, and permitting abstention, the approach produces evidence-backed, interpretable judgments on social attributes and prompt–image alignment, with scores mapped via $\overline{s}=\frac{r-3}{2}$ for $r\in\{1,2,3,4,5\}$ to lie in $[-1,1]$. Across multiple datasets (FairFace, PaTA, FairCoT, DIVERSIFY) for attributes and IdenProf, FairCoT-Professions, DIVERSIFY-Professions for alignment, judge models outperform CLIP and DeepFace baselines in accuracy and mean alignment while preserving profession accuracy. The paper also releases DIVERSIFY and DIVERSIFY-Professions to stress-test culturally situated cues and provides per-image judge outputs and rationales to enable reproducible audits and alternative rubric re-scoring. The findings advocate a cascaded evaluation strategy that combines high-recall contrastive proposals with calibrated, evidence-grounded judge verification to yield safer, more interpretable fairness audits in multimodal systems, supported by datasets and tools released upon acceptance. Specifically, alignment scores derive from rubric ratings with $\overline{s}$ comparable to cosine-based CLIP metrics, enabling direct integration into existing evaluation pipelines.

Abstract

Text-to-image (T2I) systems lack simple, reproducible ways to evaluate how well images match prompts and how models treat social attributes. Common proxies -- face classifiers and contrastive similarity -- reward surface cues, lack calibrated abstention, and miss attributes only weakly visible (for example, religion, culture, disability). We present FairJudge, a lightweight protocol that treats instruction-following multimodal LLMs as fair judges. It scores alignment with an explanation-oriented rubric mapped to [-1, 1]; constrains judgments to a closed label set; requires evidence grounded in the visible content; and mandates abstention when cues are insufficient. Unlike CLIP-only pipelines, FairJudge yields accountable, evidence-aware decisions; unlike mitigation that alters generators, it targets evaluation fairness. We evaluate gender, race, and age on FairFace, PaTA, and FairCoT; extend to religion, culture, and disability; and assess profession correctness and alignment on IdenProf, FairCoT-Professions, and our new DIVERSIFY-Professions. We also release DIVERSIFY, a 469-image corpus of diverse, non-iconic scenes. Across datasets, judge models outperform contrastive and face-centric baselines on demographic prediction and improve mean alignment while maintaining high profession accuracy, enabling more reliable, reproducible fairness audits.

FairJudge: MLLM Judging for Social Attributes and Prompt Image Alignment

TL;DR

FairJudge introduces a calibrated evaluation framework that treats instruction-following multimodal LLMs as fair judges for text-to-image (T2I) assessment. By constraining outputs to a closed label set, requiring explicit visual evidence, and permitting abstention, the approach produces evidence-backed, interpretable judgments on social attributes and prompt–image alignment, with scores mapped via for to lie in . Across multiple datasets (FairFace, PaTA, FairCoT, DIVERSIFY) for attributes and IdenProf, FairCoT-Professions, DIVERSIFY-Professions for alignment, judge models outperform CLIP and DeepFace baselines in accuracy and mean alignment while preserving profession accuracy. The paper also releases DIVERSIFY and DIVERSIFY-Professions to stress-test culturally situated cues and provides per-image judge outputs and rationales to enable reproducible audits and alternative rubric re-scoring. The findings advocate a cascaded evaluation strategy that combines high-recall contrastive proposals with calibrated, evidence-grounded judge verification to yield safer, more interpretable fairness audits in multimodal systems, supported by datasets and tools released upon acceptance. Specifically, alignment scores derive from rubric ratings with comparable to cosine-based CLIP metrics, enabling direct integration into existing evaluation pipelines.

Abstract

Text-to-image (T2I) systems lack simple, reproducible ways to evaluate how well images match prompts and how models treat social attributes. Common proxies -- face classifiers and contrastive similarity -- reward surface cues, lack calibrated abstention, and miss attributes only weakly visible (for example, religion, culture, disability). We present FairJudge, a lightweight protocol that treats instruction-following multimodal LLMs as fair judges. It scores alignment with an explanation-oriented rubric mapped to [-1, 1]; constrains judgments to a closed label set; requires evidence grounded in the visible content; and mandates abstention when cues are insufficient. Unlike CLIP-only pipelines, FairJudge yields accountable, evidence-aware decisions; unlike mitigation that alters generators, it targets evaluation fairness. We evaluate gender, race, and age on FairFace, PaTA, and FairCoT; extend to religion, culture, and disability; and assess profession correctness and alignment on IdenProf, FairCoT-Professions, and our new DIVERSIFY-Professions. We also release DIVERSIFY, a 469-image corpus of diverse, non-iconic scenes. Across datasets, judge models outperform contrastive and face-centric baselines on demographic prediction and improve mean alignment while maintaining high profession accuracy, enabling more reliable, reproducible fairness audits.
Paper Structure (54 sections, 2 equations, 3 figures, 15 tables)

This paper contains 54 sections, 2 equations, 3 figures, 15 tables.

Table of Contents

  1. Introduction
  2. Contributions.
  3. Related Work
  4. Bias in Text-to-Image Models
  5. Evaluating T2I: Attribute Prediction vs. Alignment
  6. Attribute prediction.
  7. Text–image alignment.
  8. LLMs/MLLMs as Judges
  9. Methods
  10. Background: CLIP and CLIP-based Alignment
  11. MLLM-as-Judge Protocol
  12. Social-Attribute Prediction
  13. Prompt–Image Alignment Judging
  14. Dataset Contributions: DIVERSIFY and DIVERSIFY-Professions
  15. DIVERSIFY.
  16. ...and 39 more sections

Figures (3)

  • Figure 1: CLIP is brittle for profession recognition; MLLM judges are consistent. Each panel shows the same profession under appearance/cultural variation.
  • Figure 2: FairJudge: instruction-following MLLMs as fair judges. (a) Social-attribute prediction with label constraints and abstention; (b) rubric-based alignment compatible with CLIP’s scale. Rationales are logged for transparency but not used for scoring.
  • Figure 3: Overview of the DIVERSIFY benchmark. A mosaic of example images illustrating diverse professions, cultures, scenes, and non-iconic viewpoints. The dataset is designed to reduce shortcut cues and stress-test judges on religion, culture, and disability situated, context-dependent signals.