Conformal Prediction with Paraphrase-Aware Scoring for LLM Uncertainty Quantification
Uncertainty quantification (UQ) for large language models (LLMs) aims to provide reliable measures of predictive confidence, yet current methods are often unstable under meaning-preserving perturbations. Semantically equivalent paraphrases can induce substantial variability in predictive confidence, even for methods wi...
Jiayi Xin, Evan Qiang, Zi-Han Zhu et al.
· 0 citations