Anastasios Nikolas Angelopoulos

Angelopoulos completed his Berkeley doctorate with Michael I. Jordan and Jitendra Malik, then worked with Ion Stoica as a postdoctoral scholar. His earlier research included ophthalmological technology at USC and automatic language-model evaluation at Google DeepMind. At Arena, he combines research on reliable inference with public model evaluation.[2][3]

Calibrate a set of plausible answers

The introduction he coauthored with Stephen Bates turns a fitted model into a predictor that can return several possible labels. Held-out examples measure how wrong the model is on their known labels; an adjusted quantile of those scores supplies a threshold for future prediction sets. Under exchangeability, the coverage guarantee averages over calibration and test examples. It does not promise that every individual prediction, or every subgroup, has that coverage. Python examples connect the guarantee with actual calibration and checking.[4]

Use predictions without treating them as ground truth

Prediction-Powered Inference combines a labeled experimental sample with machine-generated predictions to form valid confidence intervals for quantities such as a population mean or regression coefficient. The framework preserves statistical validity without assumptions about the prediction algorithm. Better predictions can narrow the intervals, giving researchers a way to benefit from machine learning while retaining an inferential check.[6]

Build an evaluation from pairwise preferences

The coauthored Chatbot Arena paper describes crowdsourced comparisons between model responses and statistical methods for turning votes into rankings. It studies the diversity of the submitted questions and agreement with expert raters. This measures human preference in the platform’s comparison setting; it supplies a different view of model behavior from a fixed set of questions with predetermined answers.[7]