Two-Stage MQM-Guided Framework Improves Domain-Specific Machine Translation

This framework separates translation evaluation from correction by using a retrieval-augmented LLM to generate structured MQM diagnoses before a separate LLM post-editor applies minimal edits. The approach reduces paraphrastic drift and consistently improves COMET-22 and COMETKiwi scores across seven LLMs.