Task Decomposition Does Not Improve LLM-as-a-Judge Performance

Systematic comparison shows that decomposing NLG evaluation tasks into sub-tasks provides no measurable performance gains over standard LLM-as-a-judge baselines. Reported gains in previous literature are attributed to the use of human labels in training rather than the decomposition method itself.