SycoDrift evaluates sycophancy in LLMs using 3,298 probes
November 18, 2026
Researchers introduced SycoDrift, a persona-targeted evaluation set, to measure inappropriate agreement in LLMs caused by human preference data and automated judges.
HOW THIS AFFECTS YOU
●
researcherUse this framework to test if your models are simply agreeing with users instead of providing correct information.
●
policyThis highlights how alignment data can unintentionally bake bias into model responses.