Asymmetric Conflict and Synergy in Post-training for LLM-based Multilingual Machine Translation
This work addresses efficiency in multilingual machine translation for AI and NLP applications, offering a more resource-effective alternative to scaling, though it is incremental as it builds on existing post-training methods.
The paper tackled the Curse of Multilinguality in LLM-based multilingual machine translation by analyzing linguistic conflicts and synergy during post-training, proposing a direction-aware training and group-wise model merging approach that achieved comparable performance to a baseline with 5.5x fewer pretraining tokens and 1.7x fewer parameters, resulting in only a 0.85 COMET drop on Flores-200 testsets.
The emergence of Large Language Models (LLMs) has advanced the multilingual machine translation (MMT), yet the Curse of Multilinguality (CoM) remains a major challenge. Existing work in LLM-based MMT typically mitigates this issue via scaling up training and computation budget, which raises a critical question: Is scaling up the training and computation budget truly necessary for high-quality MMT, or can a deeper understanding of CoM provide a more efficient solution? To explore this problem, we analyze the linguistic conflicts and synergy, the underlying mechanism of CoM during post-training phase. We identify an asymmetric phenomenon in linguistic conflicts and synergy: the dominance of conflicts and synergy varies in different translation directions, leading to sub-optimal adaptation in existing post-training methods. We further find that a significant bottleneck in MMT appears to lie in post-training rather than multilingual pre-training, suggesting the need for more effective adaptation strategies. Building on these new insights, we propose a direction-aware training approach, combined with group-wise model merging, to address asymmetry in linguistic conflicts and synergy explicitly. Leveraging this strategy, our method fine-tunes X-ALMA-13B-Pretrain-trained only with multilingual pre-training-achieving comparable performance to XALMA-13B (only SFT) while using only 20B pretraining tokens and 17B parameters-5.5x fewer pretraining-tokens and 1.7x fewer model size-with just 0.85 COMET drop on Flores-200 testsets of 50 languages.