What AI says about IBM FactReasoner

FactReasoner is a probabilistic factuality assessor for long-form outputs from large language models. It decomposes a generated response into atomic units, retrieves relevant external contexts for each unit, and uses probabilistic encodings of entailment and contradiction to build a joint model of facts and sources. It then computes the posterior probability that each atomic unit is supported by the retrieved contexts, enabling evaluation of factual accuracy in long-form generations.