The Attentional White Bear Effect in Transformer Language Models

Rebecca Ramnauth, Brian Scassellati

Instruction-based suppression is widely used to prevent language models from generating prohibited content, yet it remains unclear whether suppression reduces internal representati…