Prompting Toward Criticality: A Hypothesis for Fostering Creativity in Large Language Models

Abstract

Large language models can readily produce diverse and unusual outputs, yet increased variation does not by itself explain when a system becomes more capable of creative thought. This paper develops a preliminary hypothesis that prompt design may alter the effective dynamical regime of inference and, under some conditions, move a model toward a near-critical region in which small perturbations, anomalies, or competing representations have larger and more persistent effects while coherence remains sufficiently preserved. The proposal does not treat criticality as a proven property of large language models, nor does it assume that higher randomness, temperature, or output diversity is equivalent to creativity. Instead, it asks whether prompting can systematically weaken excessive restoration toward familiar representations, delay premature closure, preserve unresolved alternatives, and increase the probability that local differences influence later reasoning. Creativity is treated here in a deliberately broad but operational sense: not merely as unusual wording, but as the capacity to develop new relations, problem framings, candidate explanations, or lines of inquiry that remain consequential across subsequent inference. The paper outlines candidate prompt mechanisms, measurable perturbation-response signatures, matched-trajectory experiments, and controls that distinguish near-critical regime change from ordinary stochastic variation. It also considers epistemic turbulence as a stronger downstream possibility, to be invoked only if perturbations interact nonlinearly, persist historically, and modify later propagation pathways. The resulting framework is offered as a falsifiable research programme for testing whether prompts can do more than request creativity: whether they can alter the conditions under which creative change becomes easier to sustain.