🎉 Paper Accepted to EMNLP 2025: Introducing TaarofBench

..

We're thrilled to announce that our paper, "We politely insist: Your LLM must learn the Persian art of Taarof", has been accepted to the conference on Empirical Methods in NLP 2025, which will take place in Sazhou, China!

This paper is a collaborative effort with Prof. Ali Emami (Emory University), Nikta Gohar Sadri (Brock University) and other colleagues in Canada.

The work introduces TaarofBench, a novel benchmark designed to evaluate how well large language models (LLMs) handle taarof, a deeply embedded system of indirectness, politeness, and social negotiation in Persian discourse. While common in everyday Persian communication, taarof remains underexplored in NLP despite its linguistic and cultural complexity.

Our benchmark probes three core capabilities:

• Recognition of indirect offers or refusals

• Interpretation of the speaker’s true intent

• Generation of contextually appropriate taarof responses

Results show that even top-tier LLMs struggle with the nuances of taarof, revealing broader limitations in their ability to reason pragmatically across diverse cultural contexts.

This work is part of Zoorna Institute’s broader commitment to building culturally grounded, linguistically rich AI, especially for underrepresented languages of the Middle East and the Caucasus.