Closing the Gap at CRAC 2026: Two-Stage Adaptation for LLM-Based Multilingual Coreference Resolution

TL;DR

This paper introduces a two-stage adaptation approach using a fine-tuned multilingual LLM with adapters for improved multilingual coreference resolution, achieving top performance in the CRAC 2026 shared task.

Contribution

The authors propose a novel two-stage fine-tuning strategy with adapters and a specific mention representation, enhancing multilingual coreference resolution performance.

Findings

Achieved an average CoNLL F1 score of 74.32, ranking first in the LLM track.

Two-stage adaptation with adapters improves multilingual coreference resolution.

Design choices are effective across various languages and document types.

Abstract

We present our submission to the LLM track of the 2026 Computational Models of Reference, Anaphora and Coreference (CRAC 2026) shared task. With an average CoNLL F1 score of 74.32 on the official test set, our system ranked first in the LLM track, and third overall. Our system is based on the Gemma-3-27b model, fine-tuned using a two-stage strategy with a multilingual base adapter followed by dataset-specific adapters. We represent mention spans by their headword using an XML-inspired format with local reindexing and annotate documents iteratively. These design choices proved effective across languages, document lengths, and annotation guidelines.