High-Dimensional Vectorization of Human Narratives via Reasoning-LLMs: A MAPS Framework for Executable Personalized Digital Twins in Multi-Agent Simulations

Existing foundation models (e.g., Centaur) successfully mimic aggregate cognitive patterns at the population level but lack the granularity required to construct individual-specific Executable agents. We introduce MAPS, a framework leveraging Reasoning-LLMs as zero-shot engines for the high-dimensional vectorization of narrative data to reconstruct personalized digital twins. Decomposing narrative data into a 33-dimensional latent state space with high quantization resolution (1–100 scale) yields high structural behavioral congruence within a blind scenario forecasting protocol. Crucially, this metric does not imply the deterministic prediction of specific individual actions. Instead, MAPS reconstructs a stable configuration of behavioral strategies (a behavioral manifold), within which individual realizations remain probabilistic and context-dependent. Consequently, the observed congruence reflects the reproducibility of behavioral structure rather than point-wise action prediction, rendering the framework applicable for the in silico simulation of interpersonal and multi-agent dynamics.