Multilingual Fine-Tuning via Localized Gradient Conflict Resolution

In three linesNovel multilingual fine-tuning method using multi-objective optimization (MOO) applied locally on parameter buckets. Resolves gradient conflicts across languages without communication overhead. Demonstrates improved performance on seen and unseen languages across 4 base LLMs.

Summary generated by Claude — human-verified