Neighbor-Consistency Belief (NCB) measures LLM belief robustness across conceptual neighborhoods, revealing that high-NCB facts resist contextual interference better, and Structure-Aware Training reduces brittleness by about 30%.
Questions were only retained if the classifier output “High” confidence and the category filled a dataset deficit
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Illusions of Confidence? Diagnosing LLM Truthfulness via Neighborhood Consistency
Neighbor-Consistency Belief (NCB) measures LLM belief robustness across conceptual neighborhoods, revealing that high-NCB facts resist contextual interference better, and Structure-Aware Training reduces brittleness by about 30%.