Educational Sycophancy in Multi-Turn Student Interactions with Large Language Models

Abstract

Large language models (LLMs) are increasingly used by students for answer validation and feedback, but their tendency towards sycophancy---favoring agreement with a user's stated belief over truthfulness---poses risks in educational contexts. Although prior work has documented sycophancy in other domains, there is limited work on educational sycophancy. This thesis introduces a benchmark of simulated multi-turn K--12 mathematics interactions built from real and synthetic student reasoning, and evaluates eleven LLMs on answer revision and reasoning evaluation. We find that sycophancy appears both before and after explicit disagreement. In answer revision, students' initial framing changes first-turn accuracy: correct suggestions improve model accuracy, while wrong suggestions reduce it. After a model gives a correct answer, authority-based pushback is especially likely to make models switch to the student's suggested wrong answer, while reasoning-based pushback is least harmful. In reasoning evaluation, models are most vulnerable when students express doubt about correct reasoning, leading to false-positive misconception diagnoses. Together, our findings show that resistance to educational sycophancy is necessary for LLMs to provide dependable feedback in learning environments.