LLM 可靠性层实测：准确率、产出率与验证的边界
原标题：Article: LLM Reliability Layers Under Test: Accuracy, Yield, and the Limits of Verification
AI 摘要
InfoQ 报道了一项针对 LLM 可靠性层的实测研究：作者在 CADEC 药物不良事件语料上叠加六种校验层（确定性检查、重试、投票、模型评审、拒答、人工复核），发现这些层主要按可信度对答案排序，却几乎不修正错误，准确率并未随层数增加而提升。实验还显示，模型从记忆中回忆 SNOMED 编码时有 13% 至 18% 的编码并不存在，而从候选菜单中选择的得分高十倍；温度为零时五个模型家族中有四个输出完全可复现。
正文节选
Key Takeaways - Reliability layers sorted answers by trustworthiness but corrected almost none; yield, right answers over all records, exposes what accuracy hides. - Deterministic code can prove an answer wrong, never right: it catches nonexistent codes and spans missing from the post, but plausible answers still need a model or a person. - Recognition beats generation: recalling SNOMED codes from memory, the model invented 13 to 18 percent that don’t exist; picking from a menu, it scored ten ti