LLMs Exhibit Context-Memory Conflict Across Tasks

In a January 2026 preprint, researchers present a model-agnostic diagnostic framework that holds underlying knowledge constant while introducing controlled context–memory conflicts across tasks. Experiments on representative open-source LLMs show that degradation depends on task-specific knowledge reliance and conflict plausibility; tactics like rationale prompting increase context reliance, helping context-only tasks but harming those needing parametric memory, which biases evaluation and deployment.