davidmeikle/looped-translation-ouro-wmt24pp

Looped Translation: Fine-tuning LoopLM for NMT

First study of whether iterative latent computation helps machine translation, evaluated on WMT24++.

Quick Start

pip install -r requirements.txt
python train_ouro_2.6b.py
python train_qwen_3b.py
python eval_baseline.py
python eval_finetuned.py
python analyze_results.py

Files

• train_ouro_2.6b.py - QLoRA fine-tune Ouro-2.6B
• train_qwen_3b.py - QLoRA fine-tune Qwen2.5-3B
• eval_baseline.py - Phase 1: Baseline evaluation
• eval_finetuned.py - Phase 2: Fine-tuned evaluation
• analyze_results.py - Phase 3: Analysis

Training Data

davidmeikle/opus-translation-train-en-de-cs

Citation

@article{zhu2025scaling,
  title={Scaling Latent Reasoning via Looped Language Models},
  author={Zhu, Rui-Jie and others},
  journal={arXiv preprint arXiv:2510.25741},
  year={2025}
}