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

Disentangled World Models: Learning to Transfer Semantic Knowledge from Distracting Videos for Reinforcement Learning

arXiv 25.3 2025 54.8 method

TLDR

Disentangled World Models transfer semantic knowledge from distracting videos to RL via offline-to-online latent distillation and disentanglement constraints.

Reasoning

The paper presents a novel method for transferring semantic knowledge from distracting videos to improve RL sample efficiency, with a clear methodology combining pretraining and distillation. However, the reliance on pretrained video prediction and benchmark evaluations may limit generalizability to real-world scenarios.

Read-first score

Read-first score 54.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 38.

Recency 8%
86.7

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

Methodology quality 25%
70

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

Topical relevance 42%
54.3

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 25%
30

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 287.

Keyword Scores

world model
10
model-based reinforcement learning world model
9
world dynamics prediction
6
generative world model
5
video world model
4
world simulator
2
interactive world model
2

Deep Analysis

Innovations

  • Offline-to-online latent distillation for transferring semantic knowledge from distracting videos to world models
  • Flexible disentanglement constraints to enable cross-domain semantic knowledge transfer
  • Disentangled World Models (DisWM) framework combining pretrained action-free video prediction with online RL finetuning
  • Pretraining action-free video prediction model with disentanglement regularization to extract semantic knowledge
  • Incorporation of actions and rewards during online adaptation to strengthen disentangled representation learning

Methodology

DisWM pretrains an action-free video prediction model offline using disentanglement regularization to extract semantic knowledge from distracting videos. The pretrained model's disentanglement capability is transferred to a world model via latent distillation. During online finetuning, a disentanglement constraint is applied, and actions and rewards from environment interactions enrich data diversity to strengthen disentangled representation learning.

Key Results

Experimental results validate the superiority of the proposed approach on various benchmarks.

Resources

Paper

Tags