RLVR-World: Training World Models with Reinforcement Learning

TL;DR

RLVR-World introduces a reinforcement learning framework that directly optimizes world models for task-specific metrics, leading to significant performance improvements across language and video domains.

Contribution

It presents RLVR-World, a novel method that aligns world model training with task-specific metrics using reinforcement learning with verifiable rewards.

Findings

Improved performance on language and video world models

Effective optimization of transition prediction metrics

Versatile application across multiple domains

Abstract

World models predict state transitions in response to actions and are increasingly developed across diverse modalities. However, standard training objectives such as maximum likelihood estimation (MLE) often misalign with task-specific goals of world models, i.e., transition prediction metrics like accuracy or perceptual quality. In this paper, we present RLVR-World, a unified framework that leverages reinforcement learning with verifiable rewards (RLVR) to directly optimize world models for such metrics. Despite formulating world modeling as autoregressive prediction of tokenized sequences, RLVR-World evaluates metrics of decoded predictions as verifiable rewards. We demonstrate substantial performance gains on both language- and video-based world models across domains, including text games, web navigation, and robot manipulation. Our work indicates that, beyond recent advances in reasoning language models, RLVR offers a promising post-training paradigm for enhancing the utility of generative models more broadly. Code, datasets, models, and video samples are available at the project website: https://thuml.github.io/RLVR-World.