← 研踪 PaperTrailarXiv:2605.18549 · v2
Monitoring the Internal Monologue: Probe Trajectories Reveal Reasoning Dynamics
首次提交 2026-05-18 23:29
版本更新 2026-11-02 16:55
雷达观测 2026-11-04 23:42
摘要
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.
观测到的版本
相邻记录可能跳过中间版本；首次发现 v2 及以上也会标为版本更新。
v2 · 2026-11-02 16:55 · 最新观测版本
Monitoring the Internal Monologue: Probe Trajectories Reveal Reasoning Dynamics
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.
查看此版本解读 →查看此版本原文 ↗