Language-Centric World Models for Discrete Game Environments: A Survey of Small Model Efficiency and Planning

Abstract

Traditional world models in Reinforcement Learning (RL) often rely on high-dimensional pixel data, posing significant computational challenges for real-time decision-making. In game environments—which are fundamentally symbolic artifacts—the “ground truth” state is often discrete and structured. This paper surveys the emerging paradigm of using Small Language Models (SLMs), specifically the 8B-parameter class such as Qwen3:8B, as internal world simulators. Unlike previous surveys, we move beyond simple state-transition predic- tion P (st+1|st, at) to investigate the multi-functional roles of LLMs as Action A!ordance Generators and State/Action Validation Filters. We analyze the technical trade- o!s across the model spectrum, from 1B-parameter Edge-SLMs to 8B Small World Models (SWMs), evaluating their e”ciency in Monte Carlo Tree Search (MCTS) and proactive “What-If Analysis” (WiA). Finally, we address critical challenges in Sim-to-Real Align- ment and propose a hierarchical framework for closed-loop adaptive planning. Our findings suggest that 8B-parameter models provide an optimal equilibrium between logical fidelity and the sub-hundred-millisecond throughput required for modern game AI.

Keywords