LLM confidence judgments are dominated by a shared difficulty factor across models, with the confidence-performance link collapsing after removing agreed items, yielding no evidence for individuated metacognition.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
LLMs Show No Signs Of Individuated Metacognition
LLM confidence judgments are dominated by a shared difficulty factor across models, with the confidence-performance link collapsing after removing agreed items, yielding no evidence for individuated metacognition.