Artificial Mental Modeling by Leveraging Prescriptive Components in Large Language Models (LLMs)

Prajvi Saxena, Sabine Janzen, Wolfgang Maass

Keywords: Artificial mental models, prescriptive components, individual bias learning, cognitive modeling, LLM alignment, personalized healthcare

Abstract: Mental illness, neuro-degenerative diseases, medical rehabilitation applications critically depend on understanding individual cognitive patterns and emotional responses [Bickmore et al., 2022; Sharma et al., 2023]. However, current LLMs excel at population-level responses while fundamentally failing to model individual mental states and expectations. Existing personalization approaches rely on crude fine-tuning or prompt engineering that conflates general knowledge with individual patterns, leading to catastrophic forgetting and poor generalization. [Li et al., 2023; Kumar & Singh, 2024]. Recent work has shown that LLMs exhibit decision-making patterns comprising two distinct components, a descriptive component that reflects statistical norms from training data, and a prescriptive component that encodes implicit normative ideals, desirable, or values. [Bear & Knobe, 2017]. While prior research has focused on how these LLMs biases create problematic decision-making [Sivaprasad et al., 2025], we present a novel framework that leverages these prescriptive mechanisms that benefits individual mental modeling. We propose Artificial Mental Modeling (AMM), a personalization framework for transformer models leveraging a key insight from cognitive science: individual mental models combine descriptive knowledge (statistical knowledge of the world) with prescriptive biases (personal normative expectations) [Johnson-Laird, 1983; Norman, 2014]. Our work extends the foundational approaches of [Sivaprasad et al., 2025; Zhu et al., 2024, Wang et al., 2024] by explicitly decomposing internal representations into a descriptive subspace that captures population-level knowledge and a prescriptive subspace that encodes individual-specific goals and priors h = hdesc+hpresc. Motivated by [Voita et al., 2019; Michel et al., 2019; Clark et al., 2019] that the transformer attention heads specialize and can be selectively pruned or repurposed, AMM partitions attention heads accordingly: “descriptive heads” are frozen to preserve general world knowledge, while “prescriptive heads” are adapted with parameter efficient modules via adapters (LoRA) attached to the prescriptive heads [Hu et al., 2022]. We constrain the updates with orthogonality and sparsity regularizers to minimize interference with the descriptive subspace and mitigate catastrophic forgetting [Ravfogel et al., 2020; Kirkpatrick et al., 2017]. This separation enables targeted personalization without degrading performance across tasks requiring both generalization and individual tailoring.

Submission Number: 250