The Evolution of LLM World Models and "Word-Cultivated AI": A Multi-Paradigm Survey from Prompt Engineering to Harness Engineering, with a Framework for Verification, Pluralistic Alignment, and Reflexive Cultivation

Abstract

The notion of a "world model" — an agent's internal predictor of state transitions, rewards, and task distributions — has undergone explosive semantic broadening in the era of Large Language Models (LLMs). Across reinforcement learning, computer vision, LLM-driven social simulation, and robotics, "world model" now refers to objects that are barely comparable across communities. This survey synthesizes the rapid 2024–2026 evolution along three orthogonal threads: (i) the three engineering waves — Prompt → Context → Harness — that have reshaped how LLM-based simulation systems are built; (ii) a Levels × Laws taxonomy (L1 Humean / L2 Lewisian / L3 Lakatosian) for LLM-grounded world models; and (iii) the open problems of model drift (epistemic, identity, elaboration), pluralistic alignment integration, cognitive appraisal modeling, and security-science integration of cyber–physical–cognitive (CPC) attacks. Against this backdrop, we propose the "Word-Cultivated AI" framework: a reflexive, iteratively falsifiable methodology for cultivating LLM agents and the world models they share, grounded in five principles (Anti-Repetition, Bootstrap Expansion, Risk-Driven Termination, Falsification pre-registration, Multi-LLM cross-validation), formally specified in TLA+ and mechanically verified via the TLC model checker. The framework operationalizes Senge's Creative Tension Axis concept as a multi-axis diagnostic for "current reality vs. vision" gaps, and explicitly internalizes the methodological closure critique by requiring at least one web-grounded reviewer (e.g., Perplexity Sonar) as an outside-of-LLM-ecosystem voice. We close with a programmatic call: in a post-Mythos world where ASL-4 class capabilities are no longer scarce, world models must be cultivated through language — recursively, pluralistically, and with verifiable falsification — rather than merely engineered.

Keywords