Overview
This work separates answer changes caused by unsupported user pressure from corrections justified by new evidence. A two-turn evaluation shows that reducing the former can also impair the latter across training-time and inference-time interventions. Mechanistic analyses identify overlapping internal components, motivating interventions that preserve evidence-based updating while reducing sycophancy.
Accepted to Findings of EMNLP 2026, as reported in the paper’s arXiv record.