LLM Behavioral Consistency Improved by Fact-Heuristic-Emotion State Enforcement

A recent research paper published on arXiv presents the Cognitive Kernel Model (CKM), a prompt-level mechanism aimed at assessing and mitigating behavioral inconsistencies in large language models (LLMs) without modifying their weights. CKM compels models to categorize input into three epistemic categories: Fact (verifiable or given), Heuristic (assumed or inferred), and Emotion (priority or evaluative signal), maintaining an organized state S_t = {F_t, H_t, E_t} that is updated via a transition function. The study analyzed 26 LLMs from four different vendors in Korean-language decision-making scenarios, encompassing ambiguity, ethical dilemmas, resource distribution, and error management, resulting in 37,403 observations across four main experiments. Findings suggest that CKM can somewhat diminish instability, where LLMs provide varying responses to identical decision issues across different runs or reverse their decisions when prior answers are reintroduced as context. This method does not introduce new capabilities but ensures the tracking of information types prior to decision-making.