Don't Think of the White Bear: Ironic Negation in Transformer Models Under Cognitive Load
Abstract
Negation instructions (e.g., "do not mention X") can paradoxi- cally increase the accessibility of X in humans—a phenomenon known as ironic rebound. We investigated whether Large Lan- guage Models (LLMs) exhibit similar failures. Across nine models, we measured the probability of forbidden tokens un- der varying cognitive loads (semantic, syntactic, repetition). We found that semantic distractors induce the strongest re- bound, while repetition aids suppression. Furthermore, models with sharper polarity discrimination (distinguishing neutral from negative framings) exhibited more persistent rebound. Circuit tracing revealed that rebound is driven by a sparse set of middle- layer attention heads that amplify forbidden tokens, overpow- ering early-layer suppression. We introduce ReboundBench, a dataset of 5,000 negation prompts, to enable further study of these cognitive-like failures in AI systems.