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

WorldCompass: Reinforcement Learning for Long-Horizon World Models

arXiv 2026 64 method

TLDR

WorldCompass applies RL post-training to improve long-horizon interactive video world models via clip-level rollout, complementary rewards, and efficient RL, boosting accuracy and fidelity on WorldPlay.

Reasoning

The paper presents a novel RL framework with three clear innovations for world model training, showing strong empirical gains on a specific model. However, evaluation is limited to a single world model (WorldPlay) and lacks real-world validation, reducing generalizability.

Read-first score

Read-first score 64, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 60.

Recency 8%
100

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

Topical relevance 42%
85.7

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%
50

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

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

Frontier

Rank sensitivity

Stability: volatile; rank range: 475.

Keyword Scores

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

Deep Analysis

Innovations

  • Clip-level rollout Strategy: generating and evaluating multiple samples at a single target clip to boost rollout efficiency and provide fine-grained reward signals.
  • Complementary Reward Functions: reward functions for both interaction-following accuracy and visual quality to provide direct supervision and suppress reward-hacking.
  • Efficient RL Algorithm: negative-aware fine-tuning strategy with efficiency optimizations to enhance model capacity.

Methodology

WorldCompass is a reinforcement learning post-training framework for autoregressive video-based world models. It uses a clip-level rollout strategy for efficient sample generation and evaluation, complementary reward functions to balance interaction accuracy and visual quality, and a negative-aware fine-tuning approach with efficiency optimizations. The method is evaluated on the WorldPlay open-source world model.

Key Results

WorldCompass significantly improves interaction accuracy and visual fidelity across various scenarios when applied to the WorldPlay model.

Resources

Paper

Tags