Findings of the Fifth Shared Task on Multilingual Coreference Resolution: Expanding Datasets for Long-Range Entities
Researchers achieved first place in the LLM track of the 2026 Computational Models of Reference, Anaphora and Coreference (CRAC 2026) shared task with a system that ranked third overall. Their approach, based on the Gemma-3-27b model, employed a two-stage adaptation strategy. This method proved effective across various languages and document complexities, achieving a 74.32 CoNLL F1 score. AI
IMPACT Advances in multilingual coreference resolution could improve the performance of downstream NLP tasks like summarization and question answering.