Pairwise Comparison
Overview
Pairwise comparison studies how judgments between two items can reveal preferences, rankings, or relative quality when absolute scores are difficult to assign. Common approaches include Bradley–Terry and Elo models, low-rank statistical models, active learning to choose informative pairs, and reward models trained on preference data. These methods are used in LLM leaderboards and preference alignment, as well as recommender systems and image, audio, and video quality assessment. A central challenge is that rankings can be sensitive to judge bias, sampling noise, and non-transitive preferences, and a preferred result is not necessarily safer or objectively better.
Latest papers
81 papers1 paper in the last four weeks, against 2 the four weeks before. 0.0% of all new papers.