What AI says about M‑MAD (Multidimensional Multi‑Agent Debate)

M-MAD is a framework that uses large language models as judges to evaluate machine translation quality through multidimensional, multi-agent debates. It decomposes evaluation into distinct dimensions, conducts debates among agents for each dimension, and then a final judge synthesizes a comprehensive MT evaluation. The repository provides code, prompts, and data (with inputs from the WMT-23 Metrics Shared Task) and links to a detailed arXiv paper for methodology.