Jury Duty: Calibration and Orientation Failures in MLLM-as-a-Judge Under Cultural Ambiguity

Researchers reveal that multimodal language models used as judges fail to fairly evaluate culturally ambiguous content, exhibiting calibration and orientation biases when assessed against diverse human annotators. The study demonstrates these models systematically favor one cultural perspective while compressing their scoring scales, with implications for any AI system deployed across cultural contexts.

The research exposes a critical vulnerability in multimodal large language models (MLLMs) when tasked with making judgments across cultural boundaries. Using a benchmark of 626 paired image-prompt artifacts from U.S. and Chinese contexts, the team discovered that while individual annotator pools within each culture show strong internal agreement (α=0.86/0.74), cross-cultural evaluation diverges sharply (r=-0.12), creating a fundamental measurement problem. This matters because MLLM-as-a-Judge systems are increasingly deployed for content moderation, recommendation, and ranking—domains where cultural sensitivity directly affects user experience and fairness.

The root causes decompose into two distinct failures. First, models exhibit a positivity-floor calibration failure, compressing their use of rating scales and defaulting toward permissive evaluations. Second, they demonstrate an orientation failure, systematically aligning with one cultural norm—mechanically validating the more-lenient Chinese perspective in their experiments. Critically, persona prompting partially recovers calibration but leaves the orientation bias intact, suggesting the tilt stems from learned model weights rather than mere scale compression.

These findings carry significant implications for AI deployment at scale. When models are validated only against a single reference pool, systematic biases remain invisible. The authors' recommendation to report alignment against each reference pool separately transforms how developers should benchmark cross-cultural applications. For practitioners, this research signals that current MLLM evaluation protocols obscure rather than surface fairness issues. The invariance of model-origin bias across demonstrations suggests architectural solutions may be necessary, not just prompt engineering workarounds.

• →MLLMs systematically favor one cultural perspective while compressing scoring scales, creating dual calibration and orientation failures.
• →Within-pool annotator agreement masks cross-pool divergence, making single-reference validation insufficient for culturally heterogeneous applications.
• →Persona prompting and in-context demonstrations fail to resolve orientation bias, indicating learned model biases rather than prompt-fixable issues.
• →Model origin contributes a consistent ~0.10 MAE tilt that persists across demonstration strategies.
• →Current MLLM-as-Judge benchmarking practices obscure fairness issues when deployed across cultural contexts.