LLM Evaluation: Scoring vs. Pairwise Comparison

This paper examine Large Language Models (LLMs) used as evaluators, a concept known as "LLM-as-a-Judge," comparing two primary methods: direct scoring and pairwise comparison. The analysis indicates that pairwise comparison generally yields more reliable results and better agreement with human preferences, especially for moderately sized LLMs, due to its simpler relative judgment task. However, it also highlights that pairwise methods are susceptible to biases like positional bias and the "comparative trap." Direct scoring, while providing absolute measurements, struggles with consistency and calibration. The texts discuss strategies to enhance the reliability of both methods and note that pairwise evaluation is crucial for developing reward models in LLM alignment techniques like RLHF/RLAIF.