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

Dynamics-Aligned Latent Imagination in Contextual World Models for Zero-Shot Generalization

arXiv 25.8 2025 63.4 method, theory

TLDR

DALI infers latent context representations via self-supervised dynamics prediction for zero-shot generalization in contextual world models.

Reasoning

Strengths include a novel method with theoretical proof and strong empirical gains on cMDP benchmarks. Weaknesses are reliance on the Dreamer architecture and evaluation only in simulated environments, leaving real-world applicability unverified.

Read-first score

Read-first score 63.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 49.

Recency 8%
86.7

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

Topical relevance 42%
70

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

Methodology quality 25%
70

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

Reproducibility 25%
38

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 181.

Keyword Scores

world model
10
world dynamics prediction
9
model-based reinforcement learning world model
9
generative world model
8
world simulator
6
interactive world model
5
video world model
2

Deep Analysis

Innovations

  • Dynamics-Aligned Latent Imagination (DALI) framework for zero-shot generalization in contextual world models without requiring explicit context variables
  • Self-supervised encoder that predicts forward dynamics to infer latent context representations from agent-environment interactions
  • Theoretical proof that the dynamics-aligned encoder is essential for efficient context inference and robust generalization
  • Counterfactual consistency in latent space: perturbing a gravity-encoding dimension alters imagined rollouts in physically plausible ways

Methodology

DALI integrates within the Dreamer architecture, using a self-supervised encoder trained to predict forward dynamics from agent-environment interactions. This encoder produces latent context representations that condition the world model and policy, enabling zero-shot generalization. The framework is evaluated on contextual Markov Decision Process (cMDP) benchmarks against context-unaware and context-aware baselines.

Key Results

DALI achieves significant gains over context-unaware baselines and often surpasses context-aware baselines in extrapolation tasks, enabling zero-shot generalization to unseen contextual variations.

Limitations

  • The framework is integrated within the Dreamer architecture, which may limit applicability to other model-based RL approaches
  • Theoretical proof may rely on assumptions not fully validated in all real-world scenarios
  • Counterfactual consistency is demonstrated only for a gravity-encoding dimension; generalization to other context dimensions is not explicitly shown
  • Benchmarks are limited to cMDP environments; real-world applicability and scalability remain unverified
  • Training may require a diverse set of contexts to achieve zero-shot generalization, and performance on highly extrapolative variations is not quantified

Resources

Paper

Tags