CKM method reduces LLM decision-flip rate by 82% without retraining

Researchers Gi-Hun Lee and Joong Yull Park (arXiv preprint) propose the Cognitive Kernel Model (CKM), a prompt-level enforcement layer that makes LLMs explicitly categorize input into three epistemic roles: Fact (verifiable information), Heuristic (inferences or assumptions), and Emotion (evaluative or priority signals). CKM maintains a structured state S_t = {F_t, H_t, E_t} updated by a transition function before the model output. This doesn't change model weights or reasoning quality—only how the model tracks what kind of information it uses.

In large-scale experiments across 26 LLMs from four vendors and 37,403 observations (Korean decision scenarios: ambiguity, ethical conflict, resource allocation, error handling), CKM showed strong gains: reduced repeated-output variability (Hedges' g=1.09), cut decision-flip rates by 82% in newer models (g=1.52), and the benefit wasn't due to JSON formatting alone (value-only recomputation, g=2.24). The consistency advantage grows with sampling stochasticity (g=2.87 at temperature 0.7). A sham ablation attributes ~45% of the gain to structural scaffolding and ~55% to the fact/heuristic/emotion content. No improvement in reasoning correctness—the method only enforces behavioral consistency.