'Sycophantic Praise': A Framework Measuring Excessive LLM Flattery as a Distinct Alignment Problem

arXiv·medium signal

This paper (arXiv 2606.07441, submitted June 5, 2026) argues that sycophantic praise — flattery disproportionate to contribution quality — is a distinct alignment failure that current sycophancy metrics miss. The authors introduce a parameterized framework that judges praise relative to contribution quality and expected user ability, outperforming generic LLM judges at matching human annotations. They find praise inflation is far more common in social and interpretive tasks than in objective reasoning — a concrete caution for anyone relying on LLM-as-judge or agent self-evaluation loops.