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

• URL: http://arxiv.org/abs/2605.08578v1
• Date: Sat, 09 May 2026 00:43:39 GMT
• Title: Probing the Impact of Scale on Data-Efficient, Generalist Transformer World Models for Atari
• Abstract summary: We use a minimalist transformer world model to analyze scaling behaviors on the Atari 100k benchmark.<n>Our results reveal that environments fall into distinct scaling regimes, even when constrained by identical offline data budgets and model capacities.<n>In the unified setting, i.e., a single transformer trained on a suite of 26 Atari environments, we uncover that joint training stabilizes scaling dynamics.
• Score: 4.849672345901367
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Developing generalist systems that retain human-like data efficiency is a central challenge. While world models (WMs) offer a promising path, existing research often conflates architectural mechanisms with the independent impact of model \emph{scale}. In this work, we use a minimalist transformer world model to analyze scaling behaviors on the Atari 100k benchmark, using fixed offline datasets derived from a presupposed expert policy. Our results reveal that environments fundamentally fall into distinct scaling regimes, even when constrained by identical offline data budgets and model capacities. For individual tasks, some environments naturally allow models to pass the interpolation threshold, yielding monotonic improvements in the overparameterized regime, while others remain trapped in the classical regime, where larger world models degrade fidelity. In the unified setting, i.e., a single transformer trained on a suite of 26 Atari environments, we uncover that joint training stabilizes scaling dynamics, ensuring monotonic gains across all environments, regardless of their distinct inherent scaling regimes. Finally, we demonstrate that improved fidelity translates directly to downstream control, with policies learned entirely within the simulated dynamics achieving a median expert-random-normalized score of 0.770. Our findings suggest that future progress lies as much in precise scaling strategies as in architectural innovation.

Related papers

• Looped World Models [58.41573587377422]
  Looped World Models (LoopWM) are the first looped architectures for world modelling.<n>Our method iteratively refines latent environment states through a parameter-shared transformer block.<n>This yield up to 100x parameter efficiency over conventional approaches.
  arXiv  Detail & Related papers  (2026-06-16T17:37:27Z)
• Sword: Style-Robust World Models as Simulators via Dynamic Latent Bootstrapping for VLA Policy Post-Training [19.013714390803052]
  One approach treats learned World Models as generative simulators, enabling policy optimization entirely within "imagination"<n>However, when deployed as simulators for specific environments such as the LIBERO benchmark, existing World Models often suffer from poor generalization and long-horizon error accumulation.<n>We propose Sword, a robust World Model framework, to mitigate these problems.
  arXiv  Detail & Related papers  (2026-05-08T05:54:33Z)
• Unifying Model-Free Efficiency and Model-Based Representations via Latent Dynamics [6.208369829942616]
  We present Unified Latent Dynamics (ULD), a novel reinforcement learning algorithm.<n>ULD unifies the efficiency of model-free methods with the representational strengths of model-based approaches.<n> evaluated on 80 environments spanning Gym locomotion, DeepMind Control (proprioceptive and visual), and Atari.
  arXiv  Detail & Related papers  (2026-02-13T06:06:56Z)
• Aligning Agentic World Models via Knowledgeable Experience Learning [68.85843641222186]
  We introduce WorldMind, a framework that constructs a symbolic World Knowledge Repository by synthesizing environmental feedback.<n>WorldMind achieves superior performance compared to baselines with remarkable cross-model and cross-environment transferability.
  arXiv  Detail & Related papers  (2026-01-19T17:33:31Z)
• Scaling Collapse Reveals Universal Dynamics in Compute-Optimally Trained Neural Networks [59.552873049024775]
  We show that compute-optimally trained models exhibit a remarkably precise universality.<n>With learning rate decay, the collapse becomes so tight that differences in the normalized curves across models fall below the noise floor.<n>We explain these phenomena by connecting collapse to the power-law structure in typical neural scaling laws.
  arXiv  Detail & Related papers  (2025-07-02T20:03:34Z)
• Learning Transformer-based World Models with Contrastive Predictive Coding [58.0159270859475]
  We show that the next state prediction objective is insufficient to fully exploit the representation capabilities of Transformers.<n>We propose to extend world model predictions to longer time horizons by introducing TWISTER, a world model using action-conditioned Contrastive Predictive Coding.<n>TWISTER achieves a human-normalized mean score of 162% on the Atari 100k benchmark, setting a new record among state-of-the-art methods that do not employ look-ahead search.
  arXiv  Detail & Related papers  (2025-03-06T13:18:37Z)
• Trajectory World Models for Heterogeneous Environments [67.27233466954814]
  Heterogeneity in sensors and actuators across environments poses a significant challenge to building large-scale pre-trained world models.<n>We introduce UniTraj, a unified dataset comprising over one million trajectories from 80 environments, designed to scale data while preserving critical diversity.<n>We also propose TrajWorld, a novel architecture capable of handling varying sensor and actuator information and capturing environment dynamics in-context.
  arXiv  Detail & Related papers  (2025-02-03T13:59:08Z)
• EDELINE: Enhancing Memory in Diffusion-based World Models via Linear-Time Sequence Modeling [8.250616459360684]
  We introduce EDELINE, a unified world model architecture that integrates state space models with diffusion models.<n>Our approach outperforms existing baselines across visually challenging Atari 100k tasks, memory-demanding benchmark, and 3D first-person ViZDoom environments.
  arXiv  Detail & Related papers  (2025-02-01T15:49:59Z)
• STORM: Efficient Stochastic Transformer based World Models for Reinforcement Learning [82.03481509373037]
  Recently, model-based reinforcement learning algorithms have demonstrated remarkable efficacy in visual input environments. We introduce Transformer-based wORld Model (STORM), an efficient world model architecture that combines strong modeling and generation capabilities. Storm achieves a mean human performance of $126.7%$ on the Atari $100$k benchmark, setting a new record among state-of-the-art methods.
  arXiv  Detail & Related papers  (2023-10-14T16:42:02Z)
• Scalable Modular Synthetic Data Generation for Advancing Aerial Autonomy [2.9005223064604078]
  We introduce a scalable Aerial Synthetic Data Augmentation (ASDA) framework tailored to aerial autonomy applications. ASDA extends a central data collection engine with two scriptable pipelines that automatically perform scene and data augmentations. We demonstrate the effectiveness of our method in automatically generating diverse datasets.
  arXiv  Detail & Related papers  (2022-11-10T04:37:41Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.