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

Reinforcement Learning with Inverse Rewards for World Model Post-training

arXiv 2025 61.1 method

TLDR

Proposes RLIR, a post-training framework using inverse rewards from an Inverse Dynamics Model to improve action-following in video world models, achieving 5-10% gains.

Reasoning

The paper addresses a clear gap in world model post-training with a novel reward formulation, showing empirical gains across paradigms. However, the abstract lacks details on baselines, datasets, and limitations, and the reliance on inverse dynamics may introduce biases.

Read-first score

Read-first score 61.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 57.

Recency 8%
86.7

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

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%
50

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

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: 536.

Keyword Scores

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

Deep Analysis

Innovations

  • Proposes RLIR, a post-training framework that uses an Inverse Dynamics Model to recover input actions from generated videos, producing verifiable reward signals without large-scale preference annotations or rule-based verifiers.
  • First post-training method specifically designed to enhance action-following in video world models.
  • Maps high-dimensional video modality to low-dimensional action space for objective reward, enabling reinforcement learning optimization via Group Relative Policy Optimization.

Methodology

RLIR post-trains pre-trained world models by generating a video from an action sequence, then using an Inverse Dynamics Model to predict the action from the video. The similarity between predicted and ground-truth actions provides a reward, which is optimized with Group Relative Policy Optimization. The method is applied across autoregressive and diffusion world model paradigms.

Key Results

Achieves 5-10% gains in action-following accuracy, up to 10% improvements in visual quality, and higher human preference scores compared to existing methods.

Resources

Paper

Tags