Probing the Impact of Scale on Data-Efficient, Generalist Transformer World Models for Atari

arXiv 2026 65.1 method, theory

TLDR

Analyzes scaling behaviors of minimalist transformer world models on Atari 100k, revealing distinct scaling regimes and benefits of joint training.

Reasoning

Strengths include systematic scaling analysis and clear demonstration of joint training benefits; weaknesses are limited scope to Atari and offline expert data, lacking real-world validation.

Read-first score

Read-first score 65.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 48.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Citation impact 18%

93.3

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.93264946

Methodology quality 18%

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

Topical relevance 29%

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 18%

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

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FoundationFrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 387.

Keyword Scores

world model

model-based reinforcement learning world model

world dynamics prediction

world simulator

generative world model

video world model

interactive world model

Deep Analysis

Innovations

Isolates the independent impact of model scale from architectural mechanisms in world models
Identifies distinct scaling regimes across Atari environments, with some passing the interpolation threshold and others remaining in the classical regime
Shows that joint training on multiple environments stabilizes scaling dynamics, ensuring monotonic gains across all environments regardless of their inherent scaling regime
Demonstrates that improved world model fidelity directly translates to downstream control performance

Methodology

The study uses a minimalist transformer world model trained on fixed offline datasets derived from a presupposed expert policy on the Atari 100k benchmark. Scaling behaviors are analyzed both for individual tasks and a unified setting where a single transformer is trained on a suite of 26 Atari environments. Performance is evaluated via downstream control using policies learned entirely within the simulated dynamics, with results reported as median expert-random-normalized scores.

Key Results

Individual environments exhibit distinct scaling regimes: some show monotonic improvements in the overparameterized regime while others degrade with larger models. Joint training across 26 environments stabilizes scaling, yielding monotonic gains for all environments. The best downstream policy achieves a median expert-random-normalized score of 0.770.

Limitations

Relies on fixed offline datasets derived from a presupposed expert policy, limiting generalization to other data sources
Evaluation is confined to the Atari 100k benchmark, which may not reflect broader environments
Focuses solely on transformer world models and scale, without exploring architectural innovations or other model families

Resources

Paper