LLMs show significant performance drops on math problems in non-Western cultural contextsCultural familiarity gap of 4.88% exists when solving math problemsCultural bias persists in both prompted and zero-shot approachesGPT-4 performs better than other models but still shows cultural bias