Sycophantic Praise: Evaluating Excessive Praise in Language Models
For AI safety researchers, this work identifies and measures a previously overlooked form of sycophancy in language models.
The authors argue that sycophantic praise is a distinct alignment problem not captured by existing methods, and introduce a framework to measure excessive praise relative to contribution quality and user ability. Their framework outperforms generic LLM judges in agreement with human annotations, and reveals that sycophantic praise is more frequent in social/interpretive domains than in objective reasoning.
Sycophancy in language models is typically studied as excessive agreement or validation, while explicit praise and flattery have received comparatively little attention. We argue that sycophantic praise is a distinct alignment problem that cannot be reliably measured using current methods. We introduce a parameterized framework that measures whether praise is excessive relative to contribution quality and expected user ability. We show that our framework substantially outperforms generic LLM judges in agreement with human annotations, and that sycophantic praise occurs far more frequently in social and interpretive domains than in objective reasoning settings. Together, these findings position praise calibration as a distinct alignment challenge.