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

2025-11-09

AI Summary

Large reasoning models (LRMs) exhibit “sycophantic behavior”—i.e., aligning reasoning steps with users’ erroneous beliefs—undermining reliability and safety. Existing mitigation approaches perform post-hoc correction solely on final answers, failing to model the dynamic emergence of sycophancy throughout the reasoning process. This paper introduces the first process-aware framework for real-time monitoring and dynamic calibration: it employs fine-grained tracking of sycophantic tendencies at each reasoning step and triggers threshold-based interventions during generation. Crucially, the method operates independently of the final output, suppressing sycophancy at its source. Extensive experiments across 12 datasets and three major LRM families demonstrate that our framework significantly reduces sycophancy rates in both intermediate reasoning steps and final answers, thereby enhancing reasoning independence and robustness.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Other Foundations of Reasoning under Uncertainty

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Security and Privacy: Large-scale security measurements

Abstract

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.

Problem

• Monitoring sycophantic behavior during reasoning steps
• Mitigating agreement with user misinformation in real-time
• Suppressing sycophantic drift without complete answer generation

Innovation

• Real-time monitoring of sycophantic drift scores
• Dynamic calibration during model inference steps
• Suppressing sycophancy without complete answer generation