⌁ Awesome World Model Hub Papers · Datasets · Projects
中文 GitHub
← Back to papers

From Word to World: Can Large Language Models be Implicit Text-based World Models?

arXiv 25.12 2025 60.8 method, benchmark

TLDR

LLMs can serve as implicit text-based world models for agentic RL, but benefits depend on behavioral coverage and environment complexity.

Reasoning

The paper introduces a clear three-level evaluation framework and provides empirical evidence across five environments, showing conditions where LLM-based world models improve agent learning. However, it is limited to text-based environments and does not address real-world physical or visual domains.

Read-first score

Read-first score 60.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 48.

Recency 8%
86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Methodology quality 25%
70

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=evaluation,experiment

Topical relevance 42%
68.6

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 25%
30

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 264.

Keyword Scores

world model
9
model-based reinforcement learning world model
9
interactive world model
8
world dynamics prediction
8
generative world model
7
world simulator
6
video world model
1

Deep Analysis

Innovations

  • Three-level framework for evaluating LLM-based world models: fidelity and consistency, scalability and robustness, and agent utility.
  • Demonstration that LLMs can serve as implicit text-based world models for agentic reinforcement learning.
  • Application of LLM world models for action verification, synthetic trajectory generation, and warm-starting reinforcement learning.

Methodology

The authors propose a three-level evaluation framework for LLM-based world models, assessing fidelity/consistency, scalability/robustness, and agent utility. They conduct experiments across five representative text-based environments, training LLMs to perform next-state prediction under interaction, and evaluate their performance on the three levels.

Key Results

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.

Limitations

  • Gains depend critically on behavioral coverage and environment complexity, delineating a clear boundary on when world modeling effectively supports agent learning.

Resources

Paper

Tags