论文专业

faithfulness

忠实性。XAI 中的核心评价指标，指解释是否如实反映模型真实的内部推理过程，而非听起来合理但事后再编造的"合理化"。

English Definition

"In XAI, the degree to which an explanation accurately reflects the model's actual internal reasoning process, rather than a plausible but post-hoc rationalization."

用法说明 · 针对中文母语者

在 XAI/UX 评估里，faithfulness 与 accuracy 是两件事：accuracy 看输出对不对，faithfulness 看解释跟模型真实思考过程是否一致。中文常被误译为"准确性"。

常见错误

·把 faithfulness 跟 accuracy 混为一谈
·认为"读起来合理"的解释就是 faithful 的

真实用例 · 4 条

"a functional evaluation of faithfulness"

论文Pipeline extracted — review needed·三阶段评估中的功能层评估

"How well do MAP-X explanations demonstrate functional faithfulness by aligning with an expert-defined structure, the model's internal signal, and the ground-truth data source?"

论文Pipeline extracted — review needed·RQ1 研究问题

"generated explanations, even those using Chain-of-Thought, may appear plausible yet unfaithful to the model's true reasoning"

论文Pipeline extracted — review needed·RAG/CoT 失败模式

"MAP-X produced clinically relevant explanations with reasonable faithfulness and usability"

论文Pipeline extracted — review needed·总体结论

由 pipeline 自动采集，待人工 review