Pranay M. Mahendrakar

A Conflict Is Constructed Before It Is Measured: Seven Design Choices That Set Which Way a Language Model Bends, and Why the Context-Memory Results Do Not Compare

(c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. The literature on how language models behave when retrieved context contradicts parametric memory contains a flat contradiction, and neither of its two surveys resolves it. One body of results reports over-reliance on memorised information; another reports that mod...