UvA-MT leverages LLMs for WMT26 translation task with novel sampling and voting

Researchers from UvA-MT have detailed their approach for the WMT26 General Machine Translation shared task, focusing on the unconstrained track across 23 translation directions. Their method heavily utilizes Large Language Models (LLMs) for machine translation by employing sequential sampling with a refinement prompt to generate multiple translation candidates. These candidates are then evaluated using LLMs as pairwise judges in a round-robin voting system, a strategy that demonstrated improvement over traditional selection methods in offline WMT25 benchmark tests. AI

IMPACT Demonstrates advanced LLM application in machine translation, potentially improving translation quality and efficiency.

RANK_REASON Paper detailing participation in a machine translation shared task using LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• Hugging Face
• large-language models
• UvA-MT
• WMT25
• WMT26

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →