conference paper
Audience Behavior Modeling for Cognitive Warfare Training in Multidomain Environments
Abstract
We present an innovative architecture for Audience Behavior Modeling (ABM) that integrates Modeling & Simulation, Artificial Intelligence and Data Analytics to create a digital twin of the information environment in which it turns to be possible to make "what-if"analysis on communication strategies. Based on real world data, each agent is initialized with rich attributes - including demographic, socioeconomic, cultural, and political orientations - and evolves through feedback loops governed by system dynamics and a contextual Large Language Model (LLM). Emotional states (e.g., joy, anger, fear, etc) and traits (e.g., trust, morale, cognitive fatigue, aggressiveness) are dynamically updated in response to message exposure and social interactions. Agents possess memory, which modulates thresholds for future emotions, trust, and cognitive fatigue, allowing history-dependent behaviors. Social networks are generated using platform-calibrated, with dynamic rewiring that captures homophily and echo-chamber formation. Messages are processed through a transformer-based sentiment analyzer and fine-tuned LLM that produces stance updates and opinion vectors, enabling simulation of opinion similarity, polarization, and engagement cascades. Diffusion dynamics follow biased percolation processes with engagement decay. The model is embedded within NATO StratCom COE's Exercise SYNESIS, where real users interact with synthetic social environments; their inputs are propagated across the simulated network, and agent states are returned. This enables what-if analyses, synthetic bot generation, and strategic experimentation for cognitive and information operations in complex socio-technical ecosystems. © 2025 The Authors. Published by Elsevier B.V.
Author keywords
audience behavior modeling; cognitive warfare; modeling & simulation