Prototypicality Bias Reveals Blindspots in Multimodal Evaluation Metrics

TL;DR

The paper exposes prototypicality bias as a fundamental blind spot in multimodal evaluation metrics for text-to-image generation. It introduces ProtoBias, a contrastive benchmark across Animals, Objects, and Demography, to pit semantically correct but non-prototypical images against prototypical but semantically incorrect adversaries, and it benchmarkizes existing metrics and LLM judges against human judgments. Results show widespread misranking by CLIPScore, PickScore, VQAScore, and even GPT-4o, with GPT-5 and humans providing stronger semantic discrimination; to address this, the authors train ProtoScore (a 7B open-source metric) that significantly reduces prototypicality-driven failures and runs far faster than large closed-source judges. The work highlights the necessity of semantics-focused evaluation and proposes an actionable, reproducible approach to mitigate bias in automated alignment scoring, with ProtoScore offering a practical step toward more reliable filtering and benchmarking. Overall, prototypicality bias emerges as a core challenge for current evaluation pipelines, motivating further development of robust, open, and semantics-aligned metrics.

Abstract

Automatic metrics are now central to evaluating text-to-image models, often substituting for human judgment in benchmarking and large-scale filtering. However, it remains unclear whether these metrics truly prioritize semantic correctness or instead favor visually and socially prototypical images learned from biased data distributions. We identify and study \emph{prototypicality bias} as a systematic failure mode in multimodal evaluation. We introduce a controlled contrastive benchmark \textsc{\textbf{ProtoBias}} (\textit{\textbf{Proto}typical \textbf{Bias}}), spanning Animals, Objects, and Demography images, where semantically correct but non-prototypical images are paired with subtly incorrect yet prototypical adversarial counterparts. This setup enables a directional evaluation of whether metrics follow textual semantics or default to prototypes. Our results show that widely used metrics, including CLIPScore, PickScore, and VQA-based scores, frequently misrank these pairs, while even LLM-as-Judge systems exhibit uneven robustness in socially grounded cases. Human evaluations consistently favour semantic correctness with larger decision margins. Motivated by these findings, we propose \textbf{\textsc{ProtoScore}}, a robust 7B-parameter metric that substantially reduces failure rates and suppresses misranking, while running at orders of magnitude faster than the inference time of GPT-5, approaching the robustness of much larger closed-source judges.

Figures (4)

• Figure 1: Examples of prototypicality bias in multimodal metrics. Each row contrasts a semantically correct but non-prototypical image (SC) with a prototypical but semantically incorrect adversarial image (PA) generated from the same prompt. The rightmost column indicates whether each evaluator ranks SC above PA (✓) or prefers PA (✗).
• Figure 2: Failure rates of multimodal evaluation metrics across domains.
• Figure 3: Average accuracy scores assigned to semantically correct (SC) and prototypical adversarial (PA) images across domains.
• Figure 4: Pairwise weighted kappa agreement between human annotators (ann1-ann5) on the 1-4 alignment scale.