Multilingual large language models are increasingly embedded in daily communication, which means they must handle more than grammar and vocabulary. Figurative language such as proverbs is especially challenging because meaning depends on cultural context and conversational nuance. A new arXiv paper introduces MasalBench, a benchmark built specifically for Persian proverbs.

The benchmark targets contextual and cross-cultural understanding, probing whether models can interpret proverbs correctly when meaning shifts with situation or cultural background. The authors argue that mastering such conversational rules is essential for effective communication with users.

MasalBench aims to fill a gap in current evaluation methods, which often overlook culturally specific figurative expressions. By focusing on Persian proverbs, the benchmark offers a way to measure how well multilingual LLMs handle language that resists literal translation.