Three Regimes of Context-Parametric Conflict in LLMs

A new arXiv paper (2605.11574v1) proposes a three-regime framework to resolve contradictions in how large language models handle conflicts between training knowledge and contradicting documents. Prior studies found models either stubbornly retain trained answers (ignoring documents nearly half the time) or defer to context (approximately 96% of the time). The authors argue these contradictions arise because experiments studied three distinct processing situations without distinguishing them. Regime 1 (single-source updating) is predicted by evidence coherence; Regime 2 (competitive integration) by parametric certainty; Regime 3 (task-appropriate selection) by task knowledge requirement. The framework formalizes a distinction between parametric strength (exposure frequency) and parametric uniqueness.