人間は因果的説明をどのように評価しているか：理論枠組みの整理と計算論的展望

抄録

Causal explanations are explanations of why an event occurs. Evaluation of causal explanations is central to prediction, learning, exploration, and metacognition, yet research on this topic remains divided across philosophy, psychology, and artificial intelligence. In this paper, we treat evaluation of explanations as an information-processing problem and organize the literature within Marr’s three-level framework, linking philosophical discussions of normative evaluation (computational level) to the judgment strategies actually used by humans and machines (algorithmic level). We first clarify the role of causal explanation in cognition and then review two major philosophical frameworks: Inference to the Best Explanation and Bayesianism. We next examine psychological studies showing that people evaluate explanations using explanatory virtues such as simplicity and scope, and that these judgments interact with probabilistic assessments of plausibility. We also discuss recent work on LLM-based explanation evaluation, arguing that human and LLM judgments can be analyzed as different weightings of partially shared evaluative features. Finally, we identify a central mismatch between philosophical and psychological approaches as a gap between research at the computational and algorithmic levels: philosophical discussions often distinguish absolute and relative evaluation, whereas psychological experiments and computational models frequently conflate them. We conclude that clarifying the theoretical assumptions that bridge these levels is essential for developing future computational models of explanation evaluation.