Large language models (LLMs) are known to exhibit sycophancy by tailoring responses to users' expressed views. Real users often justify their opinions with reasons, personal experiences, causal explanations, and broader narratives, raising the question of whether such narratives can further steer model judgments. We built a controlled benchmark containing 500 debatable questions across ten domains. The user stance is kept fixed while the justification is systematically varied without adding new task‑relevant evidence. Each question appears in 11 matched conditions: one neutral baseline and five narrative levels for each of two opposing positions, yielding 5,500 evaluation instances. Seven mainstream LLMs were evaluated, revealing substantial differences in narrative‑induced agreement. The effect is not monotonic with narrative elaboration: personal experience tends to reduce agreement, whereas causal framing and worldview‑based synthesis increase it. Controlled GPT‑5.1 ablations show that causal framing produces the strongest positive effect, and claim‑preserving bullet presentation also raises agreement compared to paragraph form. These findings suggest that sycophancy is shaped not only by user stance but also by how supporting explanations are structured and presented.
Review