GLOBENUM: Benchmarking Context-Aware Numerical Formatting Reasoning Across Cultures

Numerical notation varies across cultures: 1.234 denotes one point two three four in India or the US, but one thousand two hundred and thirty-four in Belgium or Colombia. For many such expressions, the intended value cannot be recovered from the string alone; only context can resolve it. We present GLOBENUM, a benchmark that isolates these genuinely ambiguous representations and tests whether linguistic, geographic, and cultural context enables models to recover their intended value. Across 7 model families, 10 numerical formats, 12 countries, and 7 languages, using both symbolic operations and localised word problems, we find a strong preference for dot-decimal interpretations that persists despite geographic cues. Localisation further exposes a striking asymmetry: cultural context improves accuracy in dot-decimal locales while reducing it by up to 13 points in comma-decimal locales. Even with geographic and formatting cues provided, accuracy on genuinely ambiguous representations reaches only 32.78%. GLOBENUM reframes numerical reasoning as a cultural task: regional context must shape not only how numbers are written, but how their values are interpreted and computed, a distinction current models fail to make reliably.