Instruction-Following Failure

The phenomenon in which a large language model fails to generate replies in accordance with its system prompt or preset rules. It usually arises from factors such as training data, attention mechanisms or the way the user phrases a question, causing model behavior to deviate from the preset instructions.