Back to Feed
Training & Fine-Tuning / Reinforcement Learning

Improving Multilingual Translation Without Reference Data

Original: Reference-Free Post-Training of Open Large Language Models for Multilingual Machine Translation

Listen to the summary

Uses a voice available on your device

Audio options
On this page 4 sections
Related concepts 4 concepts

Key Takeaways

  • The new MiLMMT-46-12B-v1.0 model outperforms proprietary systems like GPT-5 and Gemini 3 Pro on reference-free quality metrics.
  • The method utilizes Group Relative Policy Optimization (GRPO) to improve models by training against an automatically generated reward score.
  • A reward function combining XCOMET and COMETKiwi scores, gated by language identification, removes the requirement for human-translated reference data.
  • Model interpolation between the supervised fine-tuned model and the RL-optimized model preserves learned behaviors while capturing optimization gains.
  • Reference-free post-training causes a minor reduction in spBLEU scores compared to standard supervised baselines.

Summary & Methodology Analysis

The researchers address the bottleneck of high-quality parallel data scarcity in multilingual machine translation by implementing a reference-free post-training pipeline. Starting with a supervised-finetuned (SFT, the process of refining a pre-trained model on specific tasks) MiLMMT-46-v0.1 base, the team samples candidate translations for source sentences. They compute a reward by averaging scores from two quality estimation models, XCOMET and COMETKiwi, while employing OpenLID-v3 for language identification to filter out incorrect outputs. This setup allows the system to evaluate translation quality without human-curated references.

Interactive System Flowchart

Click diagram to expand and zoom

Cross-Examination & FAQs

A deeper dive clarifying mechanics, constraints, and baseline evaluations.

Q1. What is the primary contribution of this paper?

The paper introduces a way to improve multilingual machine translation models without relying on scarce, high-quality parallel translation data.

Q2. Does the new model outperform current industry standards?

Yes, the MiLMMT-46-12B-v1.0 model outperforms recent open and proprietary systems including Google Translate, Gemini 3 Pro, and GPT-5 on reference-free scores.

Q3. Does this method require human-translated references?

No, it is a reference-free method that relies on source-side text and automated quality estimation.

Q4. What is Group Relative Policy Optimization (GRPO)?

It is an algorithm used to optimize the model policy against the reference-free reward calculated from quality estimation scores.

Q5. How is the final model produced?

The final model, MiLMMT-46-v1.0, is created by linearly interpolating the supervised fine-tuned model with the RL-optimized model.

Q6. Why is OpenLID-v3 included in the pipeline?

It is used as a gate to identify the language of the output and suppress translations that are in the wrong language.

Q7. Did the team try other methods besides RL?

Yes, they investigated on-policy distillation (OPD, a process where a smaller student model learns to mimic a larger teacher model), but it did not exceed the performance achieved by RL and checkpoint interpolation.

Q8. Are there any performance trade-offs?

Yes, reference-free post-training leads to a modest reduction in spBLEU scores compared to the supervised baselines.

Q9. What specific metrics are used to compare the models?

The paper uses reference-free quality scores and spBLEU scores, though it does not specify the exact performance gaps or hardware requirements.