Benchmarks

Reviewer3 publishes two benchmarks. On reference verification, it classifies citations as real or fabricated with 98.5% accuracy across 476 citations. On review quality, ReviewBench compares more than 145,000 human and AI review comments across computer science, social science, and life science papers, catching errors that undermine major conclusions nearly 90% of the time and ranking first in agreement with human reviewers across disciplines.