The Agent’s Hidden State Knew More Than Its Spoken Confidence

The study tests whether an agent’s internal residual-stream representations reveal eventual task success before the final answer. Latent Trajectory Dynamics summarizes how representations change across a run, while an Action Representation Probe reads states formed at action decisions. Across Bash, SQL and Python benchmarks and three open model families, both methods consistently beat surface-generation and sequence-based calibration baselines without changing prompts or sampling extra rollouts. The evidence is benchmark-specific and requires access to internal model activations, which limits applicability to closed systems.