Andreas Vlachos

This paper proposes a method to improve error prediction for LLMs by explicitly disentangling input ambiguity from standard Uncertainty Quantification signals, showing that ambiguity information significantly boosts predictive performance.