Monitoring the Internal Monologue: Probe Trajectories Reveal Reasoning Dynamics

In three linesInvestigation of LRM internal representations through probe trajectories. Authors show that continuous evolution of concept probability during reasoning predicts final behavior better than static predictions. Max-pooling achieves 95% AUROC across 4 datasets (safety, mathematics).

Summary generated by Claude — human-verified