Simple, Good, Fast: Self-Supervised World Models Free of Baggage

ICLR 25 2025 63.7 method

TLDR

SGF is a simple, fast, self-supervised world model using frame stacking and data augmentation, evaluated on Atari 100k.

Reasoning

The paper clearly identifies essential components of world models and demonstrates strong performance with a minimalist design, supported by ablation studies and quantitative comparisons. However, the evaluation is limited to the Atari 100k benchmark, and the abstract does not discuss generalization to more complex or real-world environments.

Read-first score

Read-first score 63.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 39.

Recency 8%

86.7

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

Reproducibility 25%

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

Methodology quality 25%

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

Topical relevance 42%

55.7

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

Field roles

FrontierReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 240.

Keyword Scores

world model

model-based reinforcement learning world model

world dynamics prediction

interactive world model

video world model

generative world model

world simulator

Deep Analysis

Innovations

Self-supervised representation learning without RNNs, transformers, discrete representations, or image reconstructions
Frame and action stacking to capture short-time dependencies
Data augmentation to enhance robustness against model errors

Methodology

SGF is a world model that uses self-supervised representation learning, captures short-time dependencies through frame and action stacking, and enhances robustness via data augmentation. It is evaluated on the Atari 100k benchmark with ablation studies and quantitative comparisons to established world models.

Key Results

Demonstrates good performance through quantitative comparisons on the Atari 100k benchmark.

Resources

Paper Code