Don't Think of the White Bear: Ironic Negation in Transformer Models Under Cognitive Load

2025-11-15

AI Summary

This study introduces the cognitive psychology concept of “ironic rebound” to large language model (LLM) research, investigating whether Transformer-based models exhibit target-concept reactivation under negation instructions (e.g., “do not mention X”). Using controlled experiments, polarity classification tasks, and circuit-level attention tracing, we identify two core neural mechanisms: failure of early-layer inhibition and aberrant amplification of forbidden tokens by mid-layer sparse attention heads. We propose the “polarity separation” hypothesis—that greater semantic or length disparity between instruction and forbidden term intensifies and prolongs rebound, while repeated prompting mitigates it. To systematically evaluate this phenomenon, we construct ReboundBench, a 5,000-sample benchmark, confirming the ubiquity of ironic rebound and revealing its exacerbation under long-context and semantic-interference conditions. This work establishes a novel cognitive-informed paradigm and provides empirical grounding for understanding fundamental limitations in LLM instruction following.

Technology Category

• Natural Language Processing: (Large) Language Models
• Machine Learning: Large Multimodal Models (LMMs)
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Negation instructions such as 'do not mention $X$' can paradoxically increase the accessibility of $X$ in human thought, a phenomenon known as ironic rebound. Large language models (LLMs) face the same challenge: suppressing a concept requires internally activating it, which may prime rebound instead of avoidance. We investigated this tension with two experiments. extbf{(1) Load & content}: after a negation instruction, we vary distractor text (semantic, syntactic, repetition) and measure rebound strength. extbf{(2) Polarity separation}: We test whether models distinguish neutral from negative framings of the same concept and whether this separation predicts rebound persistence. Results show that rebound consistently arises immediately after negation and intensifies with longer or semantic distractors, while repetition supports suppression. Stronger polarity separation correlates with more persistent rebound. Together, these findings, complemented by a circuit tracing analysis that identifies sparse middle-layer attention heads amplifying forbidden tokens while early layers suppress, link cognitive predictions of ironic rebound with mechanistic insights into long-context interference. To support future work, we release ReboundBench, a dataset of $5,000$ systematically varied negation prompts designed to probe rebound in LLMs.

Problem

• Investigating ironic rebound in LLMs when suppressing forbidden concepts under negation
• Examining how distractor types and cognitive load affect rebound strength in transformers
• Identifying neural mechanisms behind persistent concept activation despite suppression instructions

Innovation

• Transformer models tested under cognitive load
• Circuit tracing identifies attention heads causing rebound
• ReboundBench dataset released for future negation studies