factorized-study

This paper presents a factorised study of probe-based uncertainty estimation in LLMs, showing that raw hidden states and attention features perform well in-domain but structured features are more robust under distribution shift, and provides pretrained probes as off-the-shelf baselines.