GhazalBench: Canonical Verse Access in LLMs across Persian Ghazals and Shakespearean Sonnets
The discussion of 'GhazalBench' explores a study testing Large Language Models (LLMs) on Persian Ghazals and Shakespearean Sonnets. The core finding is that while LLMs can understand the meaning of poetry, they struggle to reproduce or recognize the exact canonical verse form. The hosts conclude that semantic understanding is separate from surface-form access, suggesting new benchmarks are needed for AI to serve as a reliable cultural steward.