A Diagnostic Framework for Auditing Reference-Free Vision-Language Metrics

ACL ARR 2025 July Submission1257 Authors

Abstract: Reference-free metrics are widely used in vision-language tasks, yet their behavior under linguistic, visual, and cultural variation remains poorly understood. We present a systematic audit of these metrics using an eight-factor diagnostic framework applied to 5,000 expert-curated MS-COCO validation images. Across dimensions such as object size, content category, syntax, named entities, spatial relations, and cultural context, we uncover consistent failure modes. Both metrics penalize captions referencing African (−5.5%, −4.8%) and Arabian (−4.9%, −5.3%) cultures, favor large-object and animal-centric scenes (plus 20 to 30 percent), and show limited sensitivity to spatial negation and word order. These findings reveal cultural and content bias, poor semantic robustness, and weak compositional understanding. We conclude with design recommendations to promote cultural fairness, scale invariance, and semantic grounding in future evaluation metrics for multimodal AI.

Paper Type: Long

Research Area: Ethics, Bias, and Fairness

Research Area Keywords: Vision Language Models, Evaluation Metrics

Contribution Types: Model analysis & interpretability

Languages Studied: English

Reassignment Request Area Chair: This is not a resubmission

Reassignment Request Reviewers: This is not a resubmission

A1 Limitations Section: This paper has a limitations section.

A2 Potential Risks: No

B Use Or Create Scientific Artifacts: Yes

B1 Cite Creators Of Artifacts: Yes

B1 Elaboration: We have cited work in reference section

B2 Discuss The License For Artifacts: N/A

B3 Artifact Use Consistent With Intended Use: Yes

B3 Elaboration: Yes we have already mentioned in sections

B4 Data Contains Personally Identifying Info Or Offensive Content: N/A

B5 Documentation Of Artifacts: Yes

B5 Elaboration: Yes section 3.2

B6 Statistics For Data: Yes

B6 Elaboration: Yes section 3.2

C Computational Experiments: Yes

C1 Model Size And Budget: No

C1 Elaboration: Not needed

C2 Experimental Setup And Hyperparameters: Yes

C2 Elaboration: Yes section 3.2-3.4

C3 Descriptive Statistics: Yes

C3 Elaboration: Yes section 3.2

C4 Parameters For Packages: Yes

C4 Elaboration: Yes section 3.2

D Human Subjects Including Annotators: No

D1 Instructions Given To Participants: N/A

D2 Recruitment And Payment: N/A

D3 Data Consent: N/A

D4 Ethics Review Board Approval: N/A

D5 Characteristics Of Annotators: N/A

E Ai Assistants In Research Or Writing: No

E1 Information About Use Of Ai Assistants: N/A

Author Submission Checklist: yes

Submission Number: 1257