mechanistic interpretability

U-Space: Revealing When and Why Language Models Generate Uncertainty

U-Space maps four uncertainty sources; U-Lens adds predictive entropy to estimate LLM reliability from one pas…