NeuroState: A Parametric Framework for Affective State Modulation in AI Systems

This record contains Version 3.1 of the NeuroState preprint. NeuroState is a parametric framework for representing affective, motivational, attentional, regulatory, and social-relational tendencies in AI systems as an N-dimensional state vector. This manuscript presents exploratory cross-model observations using two-axis positive/negative affect parameters and three-axis Valence-Arousal-Dominance (VAD) conditions across multiple large language models. The observations suggest that two-axis affective modulation produces broadly consistent qualitative trends across models, while three-axis VAD conditions capture behavioral orientation but also reveal model-specific task-assistance and instruction-following biases. The paper further proposes a six-axis NeuroState reference model, interpreted as a combination of Self State and Relation Mode, and structurally compares the framework with recent LLM emotion representation studies, including Anthropic’s 2026 work on emotion concepts in Claude Sonnet 4.5. This work does not claim that AI systems possess subjective emotions. It treats affective parameters as controllable state variables for observable behavior modulation. This manuscript is a conceptual and exploratory preprint and has not undergone peer review. A Japanese version is included as supplementary material.