The Ontology of the Alien: Escaping the Median Trap in LLM Ideation

Abstract

Large Language Models asked to "be creative" produce solutions that converge on a small number of archetypes — the Median Trap. We systematically compare eight methods for inducing structural diversity, contrasting simple prompting against three novel architectures: Semantic Tabu (accumulation of constraints), the Solution Taxonomy (a dual-agent "Studio Model"), and the Orthogonal Insight Protocol (deriving mechanisms from alternative physics). In a controlled experiment (N=196), the Studio Model exhibited emergent metacognition: the system autonomously repaired its own ontology when faced with errors and actively commissioned research into unexplored conceptual regions. Under constraint pressure, the system synthesized novel combinations that do not emerge under standard prompting, including antifragility applied to gig-worker retirement (inverting risk flows so volatility benefits the system), metric dissolution (deconstructing problem variables), and ontological accommodation (restructuring categories when data defies classification). We release the method configurations and a dataset of 196 distinctly-labeled solution archetypes, demonstrating that adversarial ontology-building forces LLMs to escape the median.

Ähnliche Arbeiten