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

ReWorld: Multi-Dimensional Reward Modeling for Embodied World Models

arXiv 26.1 2026 71.7 method, system, application

TLDR

ReWorld uses reinforcement learning to align video-based embodied world models with physical realism, task logic, and visual quality via multi-dimensional reward modeling.

Reasoning

The paper addresses a key limitation of current video world models—lack of physical fidelity and task logic—by introducing a hierarchical reward model trained on a large preference dataset and a PPO-style alignment algorithm. Strengths include a novel multi-dimensional reward approach and empirical validation; weaknesses are limited detail on specific metrics and comparisons in the abstract.

Read-first score

Read-first score 71.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 57.

Recency 8%
100

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

Topical relevance 42%
81.4

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%
80

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

Reproducibility 25%
38

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 84.

Keyword Scores

world model
10
video world model
9
model-based reinforcement learning world model
9
generative world model
8
world dynamics prediction
8
world simulator
7
interactive world model
6

Deep Analysis

Innovations

  • Multi-dimensional reward modeling for embodied world models, capturing physical realism, task completion, embodiment plausibility, and visual quality
  • Construction of a large-scale (~235K) video preference dataset for training hierarchical reward models
  • Practical alignment algorithm using a computationally efficient PPO-style method to post-train flow-based world models

Methodology

ReWorld first constructs a large-scale (~235K) video preference dataset and trains a hierarchical reward model designed to capture multi-dimensional rewards consistent with human preferences. It then proposes a practical alignment algorithm that post-trains flow-based world models using this reward through a computationally efficient PPO-style algorithm.

Key Results

ReWorld significantly improves physical fidelity, logical coherence, embodiment, and visual quality of generated rollouts, outperforming previous methods.

Limitations

  • The method is primarily evaluated on contact-rich manipulation tasks; generalization to other domains is not demonstrated.
  • The reward model relies on human preferences, which may be subjective and costly to collect at scale.

Resources

Paper

Tags