Research
5h agoApple Finds LLMs Struggle With Translation Quality
Researchers at Apple tested LLMs against human judgment using Multidimensional Quality Metrics and Error Span Annotation. The models failed to consistently match human accuracy across diverse language pairs and domains. This gap proves LLMs cannot yet replace human experts for high-precision translation audits.
The Signal
Practitioners should maintain human-in-the-loop workflows for critical MT evaluation.