Causal Understanding by LLMs: The Role of Uncertainty
Causal Understanding by LLMs: The Role of Uncertainty [43.9] 近年の論文では、LLMは因果関係分類においてほぼランダムな精度を達成している。 因果的事例への事前曝露が因果的理解を改善するか否かを検討する。 論文参考訳(メタデータ) (Wed, 24 Sep 2025 13:06:35 GMT)
LLMを因果関係を整理可能かの検証、「 uncertainty in causal tasks stems primarily from deficits in causal under- standing rather than limitations in memorization.」、「Addressing these limitations will require a shift beyond current pretraining paradigms—toward models that explicitly encode and reason over causal structures, and that are capable of expressing calibrated uncertainty when faced with ambiguity or unseen conditions.」と厳しい指摘。