Reading Calibrated Uncertainty from Language Model Trajectories

In three linesMethod to quantify uncertainty in language models by analyzing layer-wise representation trajectories. Eleven geometric features extracted from MLP updates outperform maximum softmax probability (MSP) by up to 21 AURC points, revealing where and how errors emerge across depth.

Summary generated by Claude — human-verified