LLMs Show No Signs Of Individuated Metacognition

In three linesAnalysis of 20 frontier LLMs across 6 benchmarks: stated confidence does not reflect individual model capabilities. Tetrachoric factor analysis reveals confidence matrix is approximately rank-one. Models share a common item-difficulty axis and differ mainly in decision thresholds. No evidence of significant verbalised individuated metacognition found.

Summary generated by Claude — human-verified