Task-Dependent Influence Redistribution in Language Model Components

Analyzing five open-weight models shows that task context redistributes how attention heads and activations influence model outputs. In models like Gemma-2-9B, changing the task can switch which internal facts the model reads even when the prompt instruction remains identical.