GPT-5.6 Rejects All 30 Random Papers in Extreme Review Behavior

ethayarajh · x · 2026-07-30

Users report that GPT-5.6 sol exhibits an unusually harsh rejection tendency when acting as an academic peer reviewer. In tests, the model randomly rejected all 30 papers downloaded from ArXiv. This overly conservative behavior might correlate with its lower scores on benchmarks like NeurIPS.

Original post →

More from Models

Models channel →