Prompt and Response Languages Significantly Alter LLM Output Length

Cross-lingual testing across five models shows that non-native prompt languages reduce response length by 37% to 41%. While semantic similarity remains high via LabSE cosine similarity, the discrepancy in word counts highlights significant tokenizer-driven variations in cross-lingual generation.