Presentation
Augmenting LLMs with Domain Expert Mental Models to Reduce Hallucinations
SessionWednesday Poster Session
DescriptionOur research addresses a critical limitation of large language models (LLMs): their tendency to hallucinate when faced with incomplete or missing information. While techniques such as Retrieval-Augmented Generation (RAG) and in-context learning improve factual grounding, they remain insufficient when key knowledge resides only in the minds of domain experts or when novel problems lack historical data.
To overcome this gap, we propose augmenting LLMs with Expert Mental Models (EMMs), computational representations of domain expertise derived through structured human–machine dialogue. Grounded in monotone Boolean and k-valued function theory, the approach systematically captures decision-relevant factors, organizes them hierarchically, and encodes them into prompts that guide LLM reasoning. This method transforms implicit expert knowledge into explicit, reusable structures that improve interpretability and reduce hallucinations.
A case study using Perplexity AI demonstrates the approach in the context of research proposal decision-making. Results show that while LLMs can generate useful question sets and hierarchical structures, expert refinement remains essential to address gaps and ensure decision relevance. Integrating EMMs enhances the quality of generated outputs and supports more reliable, explainable decision-making.
Our work highlights a path forward for combining human expertise with generative AI, enabling more robust, transparent, and context-aware decision support systems.
To overcome this gap, we propose augmenting LLMs with Expert Mental Models (EMMs), computational representations of domain expertise derived through structured human–machine dialogue. Grounded in monotone Boolean and k-valued function theory, the approach systematically captures decision-relevant factors, organizes them hierarchically, and encodes them into prompts that guide LLM reasoning. This method transforms implicit expert knowledge into explicit, reusable structures that improve interpretability and reduce hallucinations.
A case study using Perplexity AI demonstrates the approach in the context of research proposal decision-making. Results show that while LLMs can generate useful question sets and hierarchical structures, expert refinement remains essential to address gaps and ensure decision relevance. Integrating EMMs enhances the quality of generated outputs and supports more reliable, explainable decision-making.
Our work highlights a path forward for combining human expertise with generative AI, enabling more robust, transparent, and context-aware decision support systems.
Event Type
Poster
TimeWednesday, October 21st5:30pm - 6:30pm PDT
Location
Similar Presentations