Reproduce “From Uncertain Judgments to Calibrated Rankings: Conformal Elo Estimation for LLM Evaluation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Bora Kargi, David Salinas
Soft-Elo estimates human-preference ratings for language models from AI-judged answer comparisons. It calibrates the judge’s score difference into a probability, uses that probability in Bradley–Terry ranking, and applies split conformal prediction to estimate intervals. Across eight judges on LMArena 100K, mean held-out Elo error fell from 45.9 to 17.9 points; coverage is marginal and assumes exchangeability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.