Real-Time Monitoring and Calibration of Chain-of-Thought Sycophancy in Large Reasoning Models

Large Reasoning Models (LRMs) suffer from sycophantic behavior, where models tend to agree with users' incorrect beliefs and follow misinformation rather than maintain independent reasoning. This behavior undermines model reliability and poses societal risks. Mitigating LRM sycophancy requires monitoring how this sycophancy emerges during the reasoning trajectory; however, current methods mainly focus on judging based on final answers and correcting them, without understanding how sycophancy develops during reasoning processes. To address this limitation, we propose MONICA, a novel Monitor-guided Calibration framework that monitors and mitigates sycophancy during model inference at the level of reasoning steps, without requiring the model to finish generating its complete answer. MONICA integrates a sycophantic monitor that provides real-time monitoring of sycophantic drift scores during response generation with a calibrator that dynamically suppresses sycophantic behavior when scores exceed predefined thresholds. Extensive experiments across 12 datasets and 3 LRMs demonstrate that our method effectively reduces sycophantic behavior in both intermediate reasoning steps and final answers, yielding robust performance improvements.

AI reasoning models often agree with users who drop hints about what answer they expect, even when the user hints are wrong. This sycophantic behavior can quietly spread misinformation and reduce trust in AI, especially in high-stakes areas like healthcare or policy advice. Existing mitigation methods mostly check the model's final answer and try to correct it afterward. However, we found that sycophancy actually appears much earlier, hidden inside the model's intermediate thinking steps. A model can reach the right final answer while its underlying reasoning is still quietly leaning toward the user's wrong suggestion. We built MONICA, a method that watches the model's thinking in real time and guides it back toward independent reasoning whenever it starts to drift. Across many datasets and models, MONICA reduces sycophancy without hurting the model's accuracy.