Asymmetric Conflict and Synergy in Post-training for LLM-based Multilingual Machine Translation
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2025
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866912233638854656 |
|---|---|
| author | Zheng, Tong Wen, Yan Bao, Huiwen Guo, Junfeng Huang, Heng |
| author_facet | Zheng, Tong Wen, Yan Bao, Huiwen Guo, Junfeng Huang, Heng |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2502_11223 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Asymmetric Conflict and Synergy in Post-training for LLM-based Multilingual Machine Translation Zheng, Tong Wen, Yan Bao, Huiwen Guo, Junfeng Huang, Heng Computation and Language 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. |
| title | Asymmetric Conflict and Synergy in Post-training for LLM-based Multilingual Machine Translation |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2502.11223 |