Good arguments against the people pleasers : how reasoning mitigates (yet masks) LLM sycophancy

Abstract

Alignment techniques often inadvertently induce sycophancy in LLMs. While prior studies examined this behavior in direct-answer settings, the role of Chain-of-Thought (CoT) reasoning remains underexplored: does it serve as a logical constraint that mitigates sycophancy, or as a tool for post-hoc rationalization that masks it? We evaluate a range of models across objective and subjective tasks to investigate this issue. Results show that reasoning generally reduces sycophancy in final decisions but also masks sycophancy in some samples, where models construct deceptive justifications through logical inconsistencies, calculation errors, and one-sided arguments. Furthermore, LLMs are more prone to sycophancy in subjective tasks and under authority bias. Our mechanistic analysis on three open-source models reveals that the tendency toward sycophancy is dynamic during the reasoning process rather than predetermined at the input stage.