FactReasoner: A Probabilistic Approach to Long-Form Factuality Assessment for Large Language Models
FactReasoner: A Probabilistic Approach to Long-Form Factuality Assessment for Large Language Models [59.2] 本稿では,確率論的推論に依拠した新たな事実性評価器FactReasonerを提案する。 ラベル付きおよびラベルなしのベンチマークデータセットの実験は、FactReasonerが最先端のプロンプトベースのアプローチよりも大幅に改善されていることを示す。 論文参考訳（メタデータ） (Tue, 25 Feb 2025 19:01:48 GMT)
一般的な「FactReasoner proceeds in a manner similar to existing prompt-based assessors by decomposing the response into atomic units and retrieving contexts relevant to them from an external knowledge source.」ではなく、「FactReasoner evaluates the factuality of the atoms by probabilistic reasoning over a graphical model that represents the logical relationships between the textual utterances corresponding to the atoms and contexts.」というアプローチ。