A three-regime framework resolves contradictions in LLM context vs. parametric knowledge conflicts by distinguishing single-source updating, competitive integration, and task-appropriate selection, with empirical confirmation of certainty gradients and task effects across five models.
When not to trust language models: Investigating effectiveness of parametric and non-parametric memories
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Three Regimes of Context-Parametric Conflict: A Predictive Framework and Empirical Validation
A three-regime framework resolves contradictions in LLM context vs. parametric knowledge conflicts by distinguishing single-source updating, competitive integration, and task-appropriate selection, with empirical confirmation of certainty gradients and task effects across five models.