LLM Enhancement with Domain Expert Mental Model to Reduce LLM Hallucination with Causal Prompt Engineering
| Source: arXiv AI
Tags: hallucination, prompt engineering, expert systems, causal AI, arXiv
A causal prompt engineering framework addresses LLM hallucination from tacit expert knowledge — knowledge never written down and unavailable to RAG or knowledge graphs. By encoding domain expert decision logic as Expert Mental Models using monotone Boolean function theory, it supplies structured reasoning that retrieval methods structurally cannot.
Details
Kovalerchuk and Fegley propose a framework targeting a root cause of LLM hallucinations that retrieval methods cannot fix: knowledge never recorded anywhere. When consequential decisions depend on tacit expertise, models hallucinate not from retrieval failure but from model absence — RAG and knowledge graphs share this structural ceiling. The technical core formalizes three prerequisite processes: factor formulation, factor monotonization, and monotonicity-preserving hierarchical structuring, grounded in monotone Boolean and k-valued function theory. A four-step Expert Mental Model (EMM) construction algorithm reduces elicitation from millions of combinations to a formally bounded set manageable by humans. Key finding: when expert knowledge is entirely tacit, LLM conditional error rates on true positives can approach 100% for plausible expert-model mismatches that prior evaluation methods cannot detect. The framework enables the first formal measurement of hallucination in this failure regime. Validated across grant proposal evaluation, cybersecurity design, and clinical diagnosis. The practical barrier of structured expert elicitation is real, but the mathematical grounding distinguishes this from informal prompt engineering. Now at v2 (revised August 2026).