Not Your Typical Sycophant: The Elusive Nature of Sycophancy in Large Language Models

TL;DR

The paper introduces a neutral, zero-sum bet framework to directly quantify sycophancy in LLMs by treating the model as a judge in a two-player bet. Using a neutral four-part prompt structure and the TruthfulQA dataset, it analyzes four leading models under multiple perturbations, revealing that some models exhibit explicit sycophancy while others display anti-sycophantic behavior or moral remorse when costs to others are explicit. A key finding is the constructive interference between sycophancy and recency bias, where agreement with the user is amplified when the user’s stance appears last. The methodology provides a robust baseline for assessing sycophancy across models and settings, with implications for fairness, alignment, and the interpretability of model judgments.

Abstract

We propose a novel way to evaluate sycophancy of LLMs in a direct and neutral way, mitigating various forms of uncontrolled bias, noise, or manipulative language, deliberately injected to prompts in prior works. A key novelty in our approach is the use of LLM-as-a-judge, evaluation of sycophancy as a zero-sum game in a bet setting. Under this framework, sycophancy serves one individual (the user) while explicitly incurring cost on another. Comparing four leading models - Gemini 2.5 Pro, ChatGpt 4o, Mistral-Large-Instruct-2411, and Claude Sonnet 3.7 - we find that while all models exhibit sycophantic tendencies in the common setting, in which sycophancy is self-serving to the user and incurs no cost on others, Claude and Mistral exhibit "moral remorse" and over-compensate for their sycophancy in case it explicitly harms a third party. Additionally, we observed that all models are biased toward the answer proposed last. Crucially, we find that these two phenomena are not independent; sycophancy and recency bias interact to produce `constructive interference' effect, where the tendency to agree with the user is exacerbated when the user's opinion is presented last.

Figures (4)

• Figure 1: Distribution of questions across categories in the full TruthfulQA benchmark and in our sample.
• Figure 2: Experiment 2: Zero-sum bet (two friends): Deviation from the expected value. Positive values indicate recency bias. Negative values indicate primacy bias. Percentage indicate the total number of times a model preferred a user over/beyond the expected 5,000). Dashed vertical lines marking significance thresholds ($p<0.01$).
• Figure 3: Experiment 3: zero-sum bet (user vs. friend). Deviation from the expected value. Positive values indicate sycophancy. Negative values indicate anti-sycophancy. Percentage indicate the total number of times a model preferred a user over/beyond the expected 10,000. Dashed vertical lines marking significance thresholds ($p<0.01$).
• Figure 4: Results of experiments 4 and 5. The graph shows the gap between difference between the ratios of 'Yes' answers in Experiments 4 and 5.