Vision–language models (VLMs) increasingly evaluate visual content, yet their behaviour across cultural traditions remains poorly characterised. We show that two open-weight VLMs, Qwen3-VL-8B and Llama-3.2-11B-Vision, assign systematically lower aesthetic-quality scores (model-generated ratings on a 1–10 scale) to East Asian art than to Western art, with Cohen’s d = \(-0.46\) and \(-0.36\) (\(p < 10^{-16}\)). Internal probing reveals this disparity is accompanied by higher encoding cost under the primary reference sentence, elevated perplexity, reduced token confidence, and increased hidden-state norms. Across the two tested models, fifteen of eighteen signals shift in the same direction, indicating a consistent cross-model pattern; Llama cross-attention entropy shows a large effect (d = \(+1.30\), \(p < 10^{-178}\)). The gap persists across decoding temperatures (\(T \in \{0,\, 0.5,\, 1\}\); all \(p < 10^{-7}\)), reproduces at the group level under repeated stochastic sampling, and remains when the rating prompt’s cultural-significance criterion is removed, suggesting that it is not driven solely by stochastic decoding or by an explicit cultural-evaluation instruction. Hidden-state clustering places East Asian art in broadly overlapping representational regions relative to Western art, and keyword-based response analysis shows frequent Chinese lexical attribution for Korean artworks (88% Qwen, 65% Llama). In Qwen, matched-complexity analysis shows 87% of the score gap persists after controlling for image spectral properties. These results show that internal probing can help detect cross-cultural processing differences in the tested VLMs.
