Internal Reasoning vs. External Control: A Thermodynamic Analysis of Sycophancy in Large Language Models

Edward Y. Chang

Large Language Models exhibit sycophancy: prioritizing agreeableness over correctness. Current remedies evaluate reasoning outcomes: RLHF rewards correct answers, self-correction c…