Monitoring the Internal Monologue: Probe Trajectories Reveal Reasoning Dynamics
cs.CL, cs.CR
Submitted: 2026-05-18
Updated: 2026-11-02
License: http://creativecommons.org/licenses/by/4.0/
The gist: Large Reasoning Models (LRMs) introduce new opportunities for safety monitoring through their Chain-of-Thought (CoT) reasoning.
Terminology
Abstract
Large Reasoning Models (LRMs) introduce new opportunities for safety monitoring through their Chain-of-Thought (CoT) reasoning. However, CoT is not always faithful to the model's final output, undermining its reliability as a monitoring tool. To address this, we investigate the hidden representations of LRMs to determine whether future behavior can be predicted from prompt and CoT representations. By evaluating a probe at each generated token, we construct a probe trajectory, the continuous evolution of a concept's probability across the reasoning process. To characterize these temporal dynamics, we extract signal-processing features that capture volatility, trend, and steady-state behavior. We find that trajectory features help precisely where the static probe is weakest, adding 21.7 AUROC points on average, and improving in every such case where that probe falls below 70. We also present two methodological insights. First, template-based training data achieves near-parity with dynamically generated model responses, eliminating the need for a costly initial inference and labeling. Second, the use of a pooling operation is necessary: last-token methods achieve poor performance, whereas average and max pooling achieve 78.5% to 97.4% AUROC and yield stable probe trajectories. Across seven reasoning models from three families and six evaluation datasets, we find probe trajectories to be a promising signal for monitoring model behavior. Warning: This article contains potentially harmful content.
Sources
- Understanding intermediate layers using linear classifier probes
- Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation
- International AI Safety Report 2026
- Reasoning Models Don't Always Say What They Think
- Efficient LLM Moderation with Multi-Layer Latent Prototypes
- Training Verifiers to Solve Math Word Problems
- Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities
- Constitutional Classifiers++: Efficient Production-Grade Defenses against Universal Jailbreaks
- When Chain of Thought is Necessary, Language Models Struggle to Evade Monitors
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
- WildGuard: Open One-Stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMs
- Gaussian Error Linear Units (GELUs)
- Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety
- Building Production-Ready Probes For Gemini
- Measuring Faithfulness in Chain-of-Thought Reasoning
- DeepSeek-V3 Technical Report
- Decoupled Weight Decay Regularization
- Streaming Hallucination Detection in Long Chain-of-Thought Reasoning
- The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets
- OpenAI GPT-5 System Card
Related papers
- ALARM: Audio-Language Alignment for Reasoning Models
- ThinkSum: Probabilistic reasoning over sets using large language models
- Coherence boosting: When your pretrained language model is not paying enough attention
- From Facts to Insights: A Persona-Driven Dual Memory Framework and Dataset for Role-Playing Agents
- Improving General Role-Playing Agents via Psychology-Grounded Reasoning and Role-Aware Policy Optimization
- The Art of Asking: Multilingual Prompt Optimization for Synthetic Data