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

Scaling Laws for Pre-training Agents and World Models

arXiv 24.11 2024 50.5 theory

TLDR

Paper shows power laws in world modeling and imitation learning similar to language modeling, with coefficients influenced by tokenizer, task, architecture.

Reasoning

Strengths: Provides precise characterization of scaling laws for world models and imitation learning, extending insights from language modeling. Weaknesses: Abstract lacks details on experimental setup and real-world validation; limited to theoretical analysis without concrete results.

Read-first score

Read-first score 50.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 29.

Recency 8%
75.1

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

Methodology quality 25%
70

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

Topical relevance 42%
41.4

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%
38

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=dataset

Field roles

Methodology anchor

Rank sensitivity

Stability: volatile; rank range: 296.

Keyword Scores

world model
9
generative world model
5
world dynamics prediction
4
model-based reinforcement learning world model
4
world simulator
3
interactive world model
2
video world model
2

Deep Analysis

Innovations

  • Demonstrates that power laws similar to those in language modeling also govern world modeling and imitation learning, specifically between loss and optimal model size.
  • Reveals that the coefficients of these scaling laws are heavily influenced by the tokenizer, task, and architecture, providing nuanced guidance for optimal model and data sizing.

Methodology

The paper pre-trains embodied agents using generative learning objectives on offline datasets, modeling either agent behavior (imitation learning) or the environment (world modeling). It then characterizes scaling laws by systematically varying model parameters, dataset size, and compute, fitting power-law relationships between loss and optimal model size.

Key Results

Power laws analogous to those in language modeling are observed for both world modeling and imitation learning, but the coefficients of these laws vary significantly depending on the tokenizer, task, and architecture, affecting optimal resource allocation.

Limitations

  • The scaling law coefficients are not universal and are heavily influenced by specific choices of tokenizer, task, and architecture, limiting direct transferability across settings.
  • The analysis is based on offline datasets and generative objectives; applicability to online or other learning paradigms is not addressed.

Resources

Paper

Tags