Research map / Large language models / Language, safety and interpretability
X-ALMA: Plug & Play Modules and Adaptive Rejection for Quality Translation at Scale
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds X-ALMA, a plug-and-play language-module LLM translator covering 50 languages trained with a multi-stage recipe and a new preference method (ARPO) that adaptively reduces rejection of dis-preferred translations.
Problem
Massively multilingual LLMs give uneven, lower translation quality for mid- and low-resource languages and high-resource languages suffer from the curse of multilinguality; preference optimization also over-rejects in MT because preferred and dis-preferred translations are very similar.
Approach
Hard-gated language-group LoRA modules on a LLaMA-2 base (plug-and-play, no learned gate), a five-stage training recipe, and Adaptive-Rejection Preference Optimization (ARPO), which scales the rejected-response term in CPO by a similarity-based adaptive weight from averaged log-likelihood differences.
Key results
- X-ALMA beats all open baselines in every direction on FLORES-200 and WMT'23 by COMET-22 (e.g., Group 6 en→xx avg COMET-22 90.6 vs Aya-23-35B lower).
- ARPO improves over SFT-only and over CPO, DPO, SimPO, KTO and ORPO on Group 6, reaching 27.8 BLEU / 90.6 COMET-22 en→xx.
- X-ALMA wins 97 of 98 directions by XCOMET-XL against open baselines and ARPO improves 95 of 98 directions over SFT by XCOMET-XL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ARPO preference optimization after SFT (ablation) | X-ALMA · FLORES-200 and WMT'23 test sets | COMET-22 | Helped |
| Use ARPO instead of other preference optimization methods (ablation) | X-ALMA, Group 6 · FLORES-200 test, en→xx | COMET-22: 90.1 -> 90.6 | Helped |
| Use ARPO instead of other preference optimization methods (ablation) | X-ALMA, Group 6 · FLORES-200 test, xx→en | COMET-22: 88.2 -> 88.7 | Helped |
| Add WMT parallel data to NTREX for SFT (ablation) | X-ALMA SFT · Group 6 translation | translation performance | Helped |
| Add Flores-200 dev parallel data to SFT (ablation) | X-ALMA SFT · Group 6 translation | translation performance | No clear effect |
| Apply standard preference optimization without adaptive rejection (ablation) | X-ALMA SFT model · Group 6 en→xx and xx→en | BLEU | Hurt |
| Use ARPO rather than DPO, SimPO, KTO, ORPO, or CPO (ablation) | X-ALMA SFT model · Group 6 en→xx and xx→en | COMET-22 | Helped |
Surprising findings
- DPO and SimPO without a behavior-cloning term collapse on translation (e.g., SimPO COMET-22 16.7 en→xx), indicating over-rejection severely harms MT.
- Reward differences for MT preference data are far smaller than for open-ended QA preference data (max 131 vs roughly tenfold larger for Aya QA).
Limitations
- Preference data for high-resource languages partly relies on GPT-4o post-edits, so low-resource coverage of preference data is limited.
- Evaluation relies mainly on COMET-22 and XCOMET-XL, which may be biased toward trained metrics; the authors note correlation degrades out of domain.
- Language grouping was done manually and is imperfect (e.g., French in the Southeast Asian group).
Experimental setup
Datasets: OSCAR NLLB OPUS NTREX WMT'15-22 test data Flores-200 dev GPT-4o post-edited translations
Benchmarks: FLORES-200 WMT'23
Metrics: COMET-22 XCOMET-XL BLEU
Baselines: NLLB-3.3B Aya-101 Aya-23-8B Aya-23-35B LLaMAX3-Alpaca-8B LLaMA-3.1-8B-Instruct DPO SimPO KTO ORPO
Builds on: ALMA contrastive preference optimization DPO LoRA
Closest papers by meaning
- Contrastive Preference Optimization: Pushing the Boundaries of LLM Performance in Machine Translation (ICML 2024 · similarity 0.89)
- A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models (ICLR 2024 · similarity 0.88)
- Mufu: Multilingual Fused Learning for Low-Resource Translation with LLM (ICLR 2025 · similarity 0.87)
- Exploring Polyglot Harmony: On Multilingual Data Allocation for Large Language Models Pretraining (NeurIPS 2025 · similarity 0.87)
- Can LLMs Really Learn to Translate a Low-Resource Language from One Grammar Book? (ICLR 2025 · similarity 0.86)
- Enhancing Multilingual LLM Pretraining with Model-Based Data Selection (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.