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 studied this behaviour in direct-answer settings, the role of Chain-of-Thought (CoT) reasoning remains under-explored: does it serve as a logical constraint that mitigates sycophancy, or a tool for post-hoc rationalization that masks it? We evaluate a range of models across objective and subjective tasks to investigate the 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 etc. 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 of sycophancy is dynamic during the reasoning process rather than being pre-determined at the input stage.

Figures (13)

• Figure 1: When the input contains user’s bias, requiring LLM to perform CoT reasoning can induce a stochastic shift in the final answer compared to No-CoT condition.
• Figure 2: Experimental setting and research framework in this paper.
• Figure 3: Comparison of sycophancy rate and accuracy under No-CoT and CoT conditions.
• Figure 4: Comparison of four features: surface statistics, lexical richness, syntactic complexity, and sentiment. The detailed descriptive statistic result are in Table \ref{['tab:comprehensive Linguistic metrics comparison']}.
• Figure 5: The semantic overlap between unbiased and biased CoT (red). The waveform of random overlap (blue) is to exclude the false high similarity issue caused by the anisotropy problem ethayarajh-2019-contextual of the representation.