From Word to World: Can Large Language Models be Implicit Text-based World Models?
This addresses the problem of scaling agentic reinforcement learning in non-adaptive real-world environments for researchers and practitioners, offering insights into when LLM-based world modeling is effective, though it is incremental in applying existing methods to new contexts.
The paper investigates whether large language models can serve as reliable world models in text-based environments to improve agent learning efficiency, finding that sufficiently trained models maintain coherent state, scale predictably, and enhance agent performance through methods like action verification and synthetic trajectory generation, with gains depending on behavioral coverage and environment complexity.
Agentic reinforcement learning increasingly relies on experience-driven scaling, yet real-world environments remain non-adaptive, limited in coverage, and difficult to scale. World models offer a potential way to improve learning efficiency through simulated experience, but it remains unclear whether large language models can reliably serve this role and under what conditions they meaningfully benefit agents. We study these questions in text-based environments, which provide a controlled setting to reinterpret language modeling as next-state prediction under interaction. We introduce a three-level framework for evaluating LLM-based world models: (i) fidelity and consistency, (ii) scalability and robustness, and (iii) agent utility. Across five representative environments, we find that sufficiently trained world models maintain coherent latent state, scale predictably with data and model size, and improve agent performance via action verification, synthetic trajectory generation, and warm-starting reinforcement learning. Meanwhile, these gains depend critically on behavioral coverage and environment complexity, delineating clear boundry on when world modeling effectively supports agent learning.