CorPipe at CRAC 2026: Empty Nodes and Cross-Lingual Transfer in Multilingual Coreference Resolution

TL;DR AI
2 min readKey summary
CorPipe 26, an improved version of CorPipe 25, won the CRAC 2026 Shared Task on multilingual coreference resolution.
The shared task expanded to more datasets and languages; the system jointly predicts empty nodes, mentions, and coreference links.
CorPipe 26 outperformed other submissions in both the LLM and unconstrained tracks.
Ablations show benefits from model size, empty-node handling, and zero-shot cross-lingual transfer.
