Multilingual Fine-tuning of Large Language Models (LLMs) has achieved great advancements in machine translation. However, existing research focuses only on the traditional fine-tuning setting with a fixed set of languages, lacking dynamic adaptability to new ones. Introducing new languages requires retraining and often causes catastrophic forgetting. In this study, we propose a completely modular fine-tuning pipeline that enables dynamic language adaptation for LLMs. Instead of directly fine-tuning on all languages, our approach first trains English-centric input and output LoRA adapters for each language separately, and then merges the corresponding adapters for arbitrary-direction translation without any additional training. Experiments on 12 translation directions of four low-resource and less-supported languages show that modular fine-tuning achieves up to 86% performance of traditional multi-parallel full-parameter fine-tuning, while training only 0.1% parameters and relying solely on English-centric data without any catastrophic forgetting. Furthermore, we perform a comprehensive analysis about the merging ratio, when to merge, and the rationale for using English as a bridge language via Bayesian Optimization and logit lens.

Multilingual neural machine translation models support fine-tuning hundreds of languages simultaneously. However, fine-tuning on full parameters solely is inefficient potentially leading to negative interactions among languages. In this work, we demonstrate that the fine-tuning for a language occurs in its intrinsic language-specific subspace with a tiny fraction of entire parameters. Thus, we propose language-specific LoRA to isolate intrinsic language-specific subspaces. Furthermore, we propose architecture learning techniques and introduce a gradual pruning schedule during fine-tuning to exhaustively explore the optimal setting and the minimal intrinsic subspaces for each language, resulting in a lightweight yet effective fine-tuning procedure. The experimental results on a 12-language subset and a 30-language subset of FLORES-101 show that our methods not only outperform full-parameter fine-tuning up to 2.25 spBLEU scores but also reduce trainable parameters to 0.4% for high and medium-resource languages and 1.6% for low-resource ones.