Think Before You Link: Rarity, Reasoning, and Retrieval in Multilingual Entity Linking

Defining entity rarity through knowledge-graph structural metrics rather than pageview popularity reveals significant performance gaps, with state-of-the-art multimodal entity linking accuracy dropping by 15.4% to 39.9%. A new training-free framework improves results by using vision-language models to iteratively retrieve and reason over Wikipedia evidence.