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

Making Large Language Models into World Models with Precondition and Effect Knowledge

arXiv 24.9 2024 45.4 method

TLDR

Fine-tuning two LLMs for precondition and effect prediction enables them to function as world models, validated via human studies.

Reasoning

The paper presents a novel approach to inducing world model capabilities in LLMs through targeted fine-tuning and synthetic data, with human validation as a strength. However, it focuses narrowly on precondition and effect knowledge rather than full world dynamics, and lacks evaluation on real-world planning tasks or comparison to existing world models.

Read-first score

Read-first score 45.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 28.

Recency 8%
75.1

Uses a gentle age decay so recent papers surface without erasing older foundations. 2024

Methodology quality 25%
60

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

Topical relevance 42%
40

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

Candidate

Rank sensitivity

Stability: volatile; rank range: 278.

Keyword Scores

world model
9
world dynamics prediction
6
interactive world model
4
generative world model
3
model-based reinforcement learning world model
3
world simulator
2
video world model
1

Deep Analysis

Innovations

  • Inducing LLMs to perform world model functions of precondition prediction and effect prediction
  • Fine-tuning two separate LLMs for precondition and effect prediction respectively
  • Using synthetic data generation techniques to train the world model
  • Validating alignment of generated precondition and effect knowledge with human understanding through human-participant studies
  • Analyzing the inferred state space for supporting action chains necessary for planning

Methodology

The authors fine-tune two separate Large Language Models (LLMs) — one for precondition prediction and another for effect prediction — using synthetic data generation techniques. They then conduct human-participant studies to validate that the generated precondition and effect knowledge aligns with human understanding of world dynamics, and analyze the inferred state space for its ability to support action chains for planning.

Key Results

The precondition and effect knowledge generated by the fine-tuned LLMs aligns with human understanding of world dynamics, and the inferred state space supports the creation of action chains, a necessary property for planning.

Limitations

  • LLMs are not inherently designed to model real-world dynamics, requiring induction to perform world model functions
  • The approach relies on synthetic data generation, which may not fully capture the complexity of real-world dynamics

Resources

Paper

Tags