Deterministic External State Dynamics for Reproducible Long-Horizon LLM Agents

Abstract

Character-State Cognitive Engine (CSCE) is an external deterministic state layer for long-horizon large language model (LLM) agents. The system separates persistent character traits, dynamic latent states, policy-level behavioral biases, and language generation. Rather than allowing the LLM to serve as the canonical owner of internal state, CSCE updates state variables through explicit bounded transition functions and maps them deterministically to behavioral policy biases before language generation. The study compares this architecture with an LLM-based Character Agent baseline across controlled interventions, longitudinal trajectories, reproducibility tests, perturbation recovery, parameter robustness, and local control-theoretic analysis. The Character Agent preserves qualitative state dynamics effectively, but independent runs produce non-identical numerical latent-state trajectories. Under identical initial conditions and event sequences, CSCE produces identical state trajectories by construction. The results suggest that the main value of deterministic external state dynamics is not necessarily superior surface-level character behavior, but reproducible, inspectable, auditable, and programmatically controlled latent-state evolution for long-horizon LLM agents.