Sycophancy Suppression Can Impair Rational Updating: Anti-Sycophancy Should Preserve the Ability to Update
Published in Findings of the Association for Computational Linguistics: EMNLP 2026, 2026
When users push back, large language models often flip their answers. Such flips can come from simply aligning with the user (Unsupported-Yielding) or from genuinely useful evidence in the feedback (Rational-Updating). This work measures the two behaviors separately, finds that anti-sycophancy interventions often trade one off against the other, and shows that the two behaviors share overlapping MLP neurons and attention heads, arguing that anti-sycophancy should be treated as a selectivity problem rather than simple suppression.
Recommended citation: Huanhuan Ma, Henry Peng Zou, Chengze Li, Enze Ma, Yunyue Su, Philip S. Yu. (2026). "Sycophancy Suppression Can Impair Rational Updating: Anti-Sycophancy Should Preserve the Ability to Update." Findings of the Association for Computational Linguistics: EMNLP 2026, pp. 25499-25515.
Download Paper