Hierarchical World Models for Strategic AI Agents: Bridging Simulation and Reality through Multi-Fidelity Learning

World models allow AI agents to anticipate consequences of actions before committing to them, yet agents trained in simulation consistently exhibit a gap between tactical proficiency and strategic judgment. We formalize this phenomenon as the Strategic Simulation Gap (SSG)—the performance difference between simulated and deployed outcomes for decisions involving long-horizon reasoning under human behavioral uncertainty. We introduce Hierarchical Multi-Fidelity World Models (HMF-WM), a three-tier architecture decomposing world modeling into physical dynamics, human behavioral response prediction, and strategic outcome forecasting. A central result (Theorem 1) shows that the SSG is lower-bounded by a quantity we call the Behavioral Adaptation Coefficient, capturing how much human behavior shifts upon AI deployment—an idea drawn from the Lucas Critique in economics. Our Reality-Anchored Training (RAT) algorithm dynamically adjusts the simulation-to-real data mixture using domain verifiability as a guide. Across bug routing (v ≈ 0.8), customer escalation prediction (v ≈ 0.4), and medical triage (v ≈ 0.3), HMF-WM improves strategic accuracy by +4.3, +8.7, and +8.3 percentage points over the best baseline respectively (mean +7.1pp; all p < 0.001, paired t-test, Bonferroni-corrected), with 2.1× sample efficiency (95% CI: [1.8×, 2.4×]).