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

The paper's core contribution is not the 74.32 F1 score itself, but the demonstration that shared multilingual adapters can be decoupled from task-specific refinements without catastrophic forgetting across languages. This separation of concerns is what makes the approach portable to new languages or domains without retraining the base model.

This work sits alongside the broader shift toward parameter-efficient adaptation we've seen across recent papers. The SkillTTA work from mid-May showed how retrieval and synthesis replace fine-tuning for agent task specialization; here, adapters replace full model retraining for multilingual robustness. Both sidestep the deployment friction of parameter updates. The difference: SkillTTA operates at inference time with no parameter changes, while this CRAC system still fine-tunes but constrains it to adapter layers. The practical implication is similar though: practitioners get customization without the cost of maintaining multiple full model copies.

If the same two-stage adapter strategy maintains its F1 gains when applied to a held-out language family (e.g., Dravidian languages if the CRAC 2026 test set was Indo-European heavy), that confirms the approach generalizes. If performance drops sharply on unseen language pairs, the multilingual adapter may be overfitted to the training language distribution rather than learning a truly language-agnostic representation.