Generative Multilingual Coreference Resolution at CRAC 2026

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

Summary

A research team presented their upgraded system for generative multilingual coreference resolution at the CRAC 2026 shared task, demonstrating a substantial 15.46 percentage point uplift in CoNLL-U score. This enhanced system, which integrated the larger Gemma 3 27B IT model, joint pre-training, headword tagging, and a sliding window technique, achieved second place in the LLM track and third overall with a primary score of 73.83. The system also recorded the highest scores on two specific datasets within the competition. The paper details the implementation of more efficient training and inference strategies, alongside a comparative analysis of specialized versus general large language model approaches for coreference resolution.

Key takeaway

For NLP Engineers developing multilingual coreference resolution systems, consider integrating larger generative models like Gemma 3 27B IT. Your approach should include joint pre-training, headword tagging, and a sliding window to replicate the observed 15.46 percentage point performance uplift. Focus on efficient training and inference methods to manage the computational demands of these advanced models, potentially outperforming general LLM approaches.

Key insights

The Gemma 3 27B IT model, combined with specific techniques, significantly improves multilingual coreference resolution performance.

Principles

• Integrating larger LLMs boosts coreference scores.
• Specific architectural choices drive performance gains.
• Efficient methods are key for large model deployment.

Method

The system incorporated Gemma 3 27B IT, joint pre-training, headword tagging, efficient training/inference, and a sliding window to achieve its performance uplift.

In practice

• Consider Gemma 3 27B IT for coreference tasks.
• Apply joint pre-training for multilingual models.
• Utilize a sliding window for context management.

Topics

• Coreference Resolution
• Multilingual NLP
• Gemma 3 27B IT
• Large Language Models
• Joint Pre-training
• Efficient Inference

Best for: Research Scientist, AI Scientist, NLP 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.
• 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.
• Google's Gemma 4 AI models get 3x speed boost by predicting future tokens — Multi-Token Prediction (MTP) significantly accelerates local AI inference for Gemma 4 models via speculative decoding without quality loss.
• 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.
• Coreference Resolution in spaCy — spaCy's experimental "coref" component enables robust coreference resolution for NLP applications.

Open in AIssential →

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