Monitoring the Internal Monologue: Probe Trajectories Reveal Reasoning Dynamics

TL;DR

This paper introduces probe trajectories as a new method to monitor reasoning dynamics in large reasoning models, improving the prediction of future behavior through temporal analysis of hidden representations.

Contribution

It presents a novel approach using probe trajectories and signal-processing features to better understand and predict model behavior during reasoning tasks.

Findings

Trajectory analysis outperforms static predictions in behavior distinguishability.

Max-pooling yields up to 95% AUROC in outcome prediction.

Template-based data achieves similar results to dynamic model responses.

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. We find that future model behavior is more distinguishable when examined over the full trajectory than from a single static prediction. To characterize these temporal dynamics, we extract signal-processing features that capture volatility, trend, and steady-state behavior, significantly improving the separation of future model states. 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 choice of pooling operation is critical: average-pooling and last-token methods collapse to near-random performance, while max-pooling achieves up to 95% AUROC and yields stable probe trajectories. Using four datasets and four reasoning models across the domains of safety and mathematics, we demonstrate that trajectory features encode task-specific dynamics that improve outcome separability. These findings establish probe trajectories as a complementary framework for monitoring LRM behavior. Warning: This article contains potentially harmful content.