world model training

This work addresses the limitations of current world models, which often prioritize visual fidelity at the expense of physical plausibility and causal structure, thereby hindering their capacity for intervention, long-horizon prediction, and safety-critical decision-making. To overcome these shortcomings, the paper proposes a novel paradigm grounded in physical realism and explicit causal modeling, reframing world models as actionable simulators. The approach integrates a structured 4D interface, constraint-aware dynamics, and counterfactual reasoning mechanisms to enable precise intervention planning. Furthermore, it introduces a closed-loop evaluation framework to rigorously assess model performance. Evaluated in high-stakes domains such as medical decision-making, the method demonstrates substantial improvements in long-term robustness, intervention efficacy, and causal consistency.

This work addresses the quadratic computational and memory overhead of conventional Transformers with respect to context length, which hinders efficient predictive understanding and controllable simulation in time-series modeling. To overcome this limitation, the authors propose World Machine—a generative world model grounded in a structured latent state mechanism that circumvents the quadratic complexity of standard Transformers and dynamically adapts to both observation volume and context length. The approach integrates a modified Transformer architecture, latent state representation learning, and a tailored training protocol. Evaluated on a custom Toy1D synthetic dataset, World Machine demonstrates superior modeling capacity compared to baseline Transformers, with ablation studies confirming the contribution of each component to overall performance gains.

This work addresses the limitations of traditional world models that rely on recursive multi-step prediction, which often suffer from error accumulation and degraded long-horizon forecasting accuracy. The authors propose a novel end-to-end, non-recursive Direct Prediction World Model (DPWM) that compresses action sequences of arbitrary length into a single embedding and directly predicts the final state in a single forward pass. By avoiding recursive unrolling, DPWM mitigates error propagation and training instability while enabling gradient flow over extended horizons. Evaluated on continuous control and pixel-based benchmarks, the method consistently outperforms existing recursive models, with performance gains becoming more pronounced as the prediction horizon increases—demonstrating the critical role of terminal-state prediction as an effective objective for long-horizon tasks.

Existing pre-trained video generation models rely on static prompts (e.g., text or images), limiting their ability to model interactive, dynamic scenes. To address this, we propose the Dynamic World Simulation (DWS) framework, which transforms video generation models into interactive world simulators. DWS introduces a lightweight, universal action-conditioning module that drives scene evolution according to given action trajectories; a motion-augmented loss that explicitly optimizes dynamic consistency—rather than pixel-level fidelity; and a priority imagination sampling strategy to enhance long-horizon temporal controllability. The framework is architecture-agnostic, supporting both diffusion models and autoregressive Transformers. Experiments demonstrate that DWS significantly improves action controllability and dynamic coherence in game and robotics simulation scenarios. Moreover, when applied to downstream model-predictive control tasks, DWS achieves state-of-the-art sample efficiency.

To address catastrophic forgetting in continual physical dynamics modeling under nonstationary visual environments, this paper proposes a forgetting-free Mixture-of-Gaussians World Model (MG-WM). MG-WM integrates predictive experience replay with an exploration-conservative dual-value estimation mechanism, enabling stable visual dynamics modeling and policy optimization amid environmental evolution. It is the first method to unify three critical properties—environmental simulation continuity, predictive guidance of experience replay, and exploration-conservative balance in behavioral policies—within a model-based reinforcement learning (MBRL) framework. Evaluated on the DeepMind Control and Meta-World continual control benchmarks, MG-WM significantly outperforms state-of-the-art methods. Moreover, on evolving-domain video prediction tasks, it effectively mitigates spatiotemporal dynamics forgetting, demonstrating robust adaptation to dynamic distribution shifts.

This work addresses the limitations of traditional world models, which are confined to predicting physical states and thus fail to provide the diverse, actionable feedback necessary for agents’ continual learning. To overcome this, the paper proposes Agent-Centric Interactive World Proxies that extend world modeling beyond state transitions to multimodal information transfer, encompassing execution outcomes, experience retrieval, skill invocation, and validation signals. The framework innovatively decomposes the world proxy into six functional modules—dynamics, spatial, execution, memory, skill, and reward—and integrates them within a hierarchical augmentation architecture. This design operates across three levels: inference guidance, training optimization, and co-evolution of proxies and agents. Empirical results demonstrate that this paradigm substantially enhances agents’ planning capabilities, learning efficiency, and potential for sustained evolution.

Existing latent world models rely on single-step prediction, leading to error accumulation during recursive rollout in long-horizon planning and suffering from a mismatch between their training objective and the actual planning task. This work proposes the Variable-Length World Model (VLWM), which introduces, for the first time, a mechanism for predicting future latent states based on variable-length action sequences. VLWM employs a curriculum learning strategy that progressively optimizes the model from short- to long-horizon predictions and is accompanied by a tailored latent-space planning algorithm. Evaluated across multiple long-horizon control tasks, VLWM outperforms the current state-of-the-art method, LeWM, by an average of 13%, with particularly pronounced gains in tasks requiring extended planning horizons.

This work addresses the high complexity and resource demands of existing video world models, which lack lightweight and reproducible baselines. The authors propose MiniWorld, an end-to-end trainable streaming video world model based on a block-causal video diffusion Transformer. It operates in the latent space of a pretrained Video VAE, trained via flow matching with a novel block-wise non-decreasing noise schedule and a two-stage continual training strategy to enhance temporal modeling. MiniWorld provides the first fully reproducible, lightweight from-scratch training protocol, enabling completion within days on a single 8-GPU server. Efficient streaming inference is achieved through rolling KV caching and an asynchronous denoising pipeline. The release of code and models significantly lowers the barrier to entry for future research in this domain.

This work addresses the challenge of unifying prediction, planning, and irreversibility within world models. It formulates prediction as a probability measure over future trajectories and, under a local Markov assumption, employs the Onsager–Machlup action functional to decompose latent dynamics into reversible and irreversible components in path space. The authors introduce rollout-based entropy production as an operational measure of irreversibility. Through path-integral analysis, attention mechanism inspection, and small-scale model experiments, they find that attention asymmetry emerges in response to increasing data irreversibility. While symmetrization interventions suppress entropy production, they selectively impair long-horizon prediction of irreversible processes yet preserve the model’s capacity to capture relaxation dynamics.

This work addresses a critical limitation in existing latent-variable world models, which rely on average prediction error over training data for training and selection—a metric that fails to reflect actual controller performance due to a mismatch between the evaluation distribution and the distribution queried by the planner. The authors propose instead to center model assessment on the discrepancy between predicted and true costs over states reachable by the planner. They establish, for the first time, a rigorous theoretical link between this discrepancy and control suboptimality, proving it provides a valid upper bound on performance loss, whereas conventional prediction errors neither bound nor track performance. Leveraging control theory, spectral analysis, and non-normal operator theory, they decompose the discrepancy into an intrinsic manifold residual and an off-manifold divergence term, and introduce a fidelity score to quantify alignment of the planner’s reachable distribution. Experiments on synthetic systems and model predictive control confirm that the proposed metric reliably tracks control performance, while single-step prediction error shows virtually no correlation.