A framework for analyzing AI leaderboard stability and manipulation
The paper’s arXiv description highlights how leaderboards such as LMArena turn human preferences between pairs of language models into rankings.
TLDR
A post shares the arXiv paper “A Unified Perturbation Framework for Analyzing Leaderboard Stability and Manipulation.” Its arXiv description frames these leaderboards as central to benchmarking large language models, citing LMArena and its use of pairwise human preferences to produce rankings.
Combined views
3.2K
1 Source, first seen ago
