ShaQ: Localizing Input Uncertainty in LLMs via Shapley Values

A new framework called Shapley-based input uncertainty Quantification (ShaQ) has been introduced by researchers to address span-level attribution of uncertainty induced by inputs in large language models (LLMs). Existing techniques focus on output levels and do not differentiate between gaps in knowledge and ambiguities in input. ShaQ offers detailed, practical insights into which segments of the input contribute to uncertainty. This research has been made available on arXiv under the identifier 2605.28170.