What is pairwise ranking evaluation in model comparison?

Pairwise ranking evaluates models by comparing outputs two at a time.

Instead of absolute scoring, pairwise evaluation asks which of two model outputs is better. Aggregated over many comparisons, it yields a global ranking using methods like Bradley-Terry or Elo systems. This reduces calibration issues in human or LLM judgment and improves consistency in subjective tasks.