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

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, short

Summary

A novel system for LLM-based multilingual coreference resolution secured first place in the LLM track and third overall at the 2026 Computational Models of Reference, Anaphora and Coreference (CRAC 2026) shared task. This high-performing system, which uses the Gemma-3-27b model, achieved an impressive average CoNLL F1 score of 74.32 on the official test set. Its core methodology involves a sophisticated two-stage fine-tuning strategy, beginning with a multilingual base adapter and subsequently applying dataset-specific adapters. Further enhancing its capabilities, the system represents mention spans by their headword using an XML-inspired format with local reindexing and employs an iterative process for document annotation. These specific design choices collectively proved highly effective across diverse languages, varying document lengths, and different annotation guidelines.

Key takeaway

For NLP Engineers developing multilingual coreference resolution systems, this research indicates that a two-stage adapter fine-tuning approach on models like Gemma-3-27b can significantly boost performance. You should consider implementing a multilingual base adapter followed by dataset-specific adapters to achieve superior CoNLL F1 scores. Additionally, integrating XML-inspired headword representation and iterative document annotation into your workflow could enhance effectiveness across diverse linguistic contexts and document structures.

Key insights

Two-stage adapter fine-tuning on Gemma-3-27b significantly improves multilingual coreference resolution performance.

Principles

• Adapter-based fine-tuning enhances LLM multilingual performance.
• Iterative annotation improves document processing.
• Headword-based XML representation is effective for mention spans.

Method

The system fine-tunes Gemma-3-27b using a two-stage strategy: a multilingual base adapter followed by dataset-specific adapters. It represents mention spans by headwords in an XML-inspired format with local reindexing and annotates documents iteratively.

In practice

• Implement two-stage adapter fine-tuning for multilingual tasks.
• Use XML-inspired headword representation for coreference.
• Adopt iterative document annotation workflows.

Topics

• LLM Fine-tuning
• Coreference Resolution
• Multilingual NLP
• Gemma-3-27b
• Adapter Networks
• Shared Task Performance

Best for: Research Scientist, AI Scientist, NLP Engineer, Machine Learning Engineer

Related on AIssential

• Closing the Gap at CRAC 2026: Two-Stage Adaptation for LLM-Based Multilingual Coreference Resolution — Two-stage fine-tuning and optimized input formats enable LLMs to achieve state-of-the-art multilingual coreference resolution.
• Multilingual Coreference Resolution via Cycle-Consistent Machine Translation — Coreference resolution in low-resource languages can be significantly improved by cycle-consistent machine translation data augmentation.
• A Larger Annotated Corpus of Portuguese Coreference — LLMs can annotate Portuguese coreference, but a new LLM-based system did not outperform existing rule-based methods.
• Plug-and-Adapt: Multimodal Coreference Resolution at First Sight with a Pretrained Alignment Model — Plug-and-Adapt repurposes a pre-trained vision-language alignment model for zero-shot multimodal coreference resolution, outperforming existing methods.
• CorPipe at CRAC 2026: Empty Nodes and Cross-Lingual Transfer in Multilingual Coreference Resolution — CorPipe 26 integrates empty node prediction with mention and coreference resolution for improved multilingual performance.
• System Report for CCL25-Eval Task 5: New Dataset and LoRA-Fine-Tuned Qwen2.5 — Domain-specific LoRA fine-tuning on Qwen2.5-14B with CCPoetry-49K significantly enhances classical Chinese poetry understanding and translation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.