MLLM-as-Judge shows cultural bias: calibration and orientation failures revealed

Researchers from multiple institutions have released a paper examining how multimodal large language models (MLLMs) perform as judges when evaluating culturally ambiguous content. They created VOIR DIRE, a benchmark of 626 culturally paired image–prompt artifacts spanning food, fashion, and architecture from U.S. and mainland Chinese contexts. Annotator pools were internally reliable (α = 0.86 and 0.74) but cross-pool evaluations diverged significantly (Q1 r = -0.12). Testing six MLLMs, the team identified two distinct failure modes: a positivity-floor calibration failure where models compressed their rating scales (avoiding low scores) and an orientation failure where models defaulted to one cultural norm. On contested items that split the two pools, this floor effect mechanically validated the more-permissive Chinese reading.

Efforts to mitigate these biases had limited success. Persona prompting (asking the model to adopt a specific cultural perspective) partially recovered calibration but the orientation residual survived—evidence the tilt is not reducible to scale compression. Using reference-pool in-context demonstrations actually deepened the orientation residual and inflated high-end scores rather than restoring low-end use. Model origin (e.g., a model trained primarily on U.S. data) added a small additive tilt of about 0.10 MAE that remained largely invariant under demonstration. The authors recommend that in practice, alignment should be reported against each cultural reference pool separately, and cross-pool divergence should be treated as a property of the judge itself rather than an error to be eliminated.