Beyond the Sycophancy Score: How Task, Model, and Pressure Shape LLM Yielding.

arXiv:2610.08840v1 Announce Type: cross Abstract: Large language models (LLMs) often abandon a correct answer, or endorse a user's position, once the user pushes back. This behavior, called sycophancy, is usually reported as a single rate per model, which says