Large language models can do more than generate text: with the right fine-tuning, relatively small models can serve as simulators of an environment, predicting how the world changes after each action an AI agent takes. That is the finding of a paper posted to the arXiv preprint server on December 21, 2025, by researchers from China’s Southern University of Science and Technology, Princeton University, Microsoft Research and other institutions.
The paper, titled “From Word to World: Can Large Language Models be Implicit Text-based World Models?”, reframes language modeling as next-state prediction. Instead of asking a model to predict the next word, the researchers ask it to predict the next state of a text-described environment given the current state and an agent’s action.
What the researchers tested
The team evaluated models across five text-based environments of varying structure, including household task simulations, laboratory-style science experiments, narrative puzzle games, an e-commerce shopping environment and a tool-use benchmark. They propose a three-level framework for judging a language model as a world model: whether its predictions are accurate and internally consistent, whether performance scales with data and model size, and whether the simulated experience actually helps an agent learn.
According to the paper and summaries of it, open models in the 7 to 8 billion parameter range, including Alibaba’s Qwen2.5-7B and Meta’s Llama 3.1-8B, reached roughly 99 percent state-prediction accuracy on the household environment after fine-tuning on interaction logs, and close to that on the science environment. In structured domains, the authors report that plans formed inside the simulator succeeded in the real environment at nearly the same rate, with consistency above 90 percent over multi-step interactions.
Performance was weaker in open-ended settings. In the narrative puzzle environment, where the state space is larger and less predictable, accuracy fell to around 70 percent. The authors write that the gains depend critically on how well the training data covers the agent’s behavior and on the complexity of the environment, and they describe their results as drawing a boundary around where world modeling helps.
Why it matters for agent training
The motivation is practical. Training agents through reinforcement learning requires large amounts of interaction, and real environments, whether websites, software tools or physical settings, are slow, expensive and hard to scale. A language model that can stand in for the environment lets an agent practice in a fast, controllable simulation. The paper reports that such simulators improved agent performance in three ways: verifying candidate actions before they are taken, generating synthetic training trajectories, and warm-starting reinforcement learning.
The work fits a broader argument in the field, articulated by Google DeepMind researchers David Silver and Richard Sutton in a 2025 essay titled “Welcome to the Era of Experience,” that future AI systems will learn mainly from interacting with environments rather than from static, human-generated datasets. The new paper does not settle whether language models “understand” the world in any deep sense, but it does provide evidence that, within well-defined domains, they can learn to model its rules and consequences well enough to be useful for training other systems.
