Atomic and Holistic LLM Judges for Reference-Grounded Support Labels: A Prompt-Controlled Comparison
The study compared two methods for classifying reference support: an atomic judge that breaks answers into individual claims versus a holistic judge using a detailed rubric. On completeness-heavy tasks like ASQA, the holistic approach was competitive or superior, especially for finding missing information. Atomic decomposition did not offer significant accuracy gains and used more tokens.