Prompt Perturbation for Reliable LLM Evaluation over Comparison Graphs

In three linesMethod to evaluate LLMs via pairwise comparisons by resolving intransitivity (cycles A≻B≻C≻A). Prompt perturbation framework generates prompt variants, identifies structural inconsistencies in comparison graphs, then applies filtered ranking methods to stabilize leaderboards.

Summary generated by Claude — human-verified