2026-10-11
Apple study finds LLMs and humans both struggle to grade machine translation
Apple researchers report that large language models, tested against human judgment on the MQM and ESA quality-annotation schemes across varied language pairs and domains, show mixed reliability, per machinelearning.apple.com.
What else the source says
- Apple researchers tested LLMs against human judgment for MT quality evaluation
- Two annotation schemes compared: MQM and ESA
- Evaluation covered multiple language pairs, domains and granularity levels
Why it matters
Apple's finding suggests LLM translation grading still needs human oversight.
Source. machinelearning.apple.com
TRACE
- First seen Oct 12, 2026, 3:31 PM PT
- Published Oct 12, 2026, 4:34 PM PT, 1.1 h after first seen
- Source machinelearning.apple.com (Company source)