How AI describes IBM FactReasoner

FactReasoner is a probabilistic factuality assessment framework for Large Language Models that decomposes responses into atomic claims and verifies them against external knowledge sources. It provides calibrated probability estimates for each claim, enabling nuanced uncertainty quantification through a five-stage pipeline including atomic decomposition, decontextualization, context retrieval, NLI-based verification, and probabilistic reasoning.