RLVR-World: Training World Models with Reinforcement Learning
This work addresses the problem of improving world model utility for researchers and practitioners by offering a post-training paradigm, though it is incremental as it builds on existing reinforcement learning and world modeling techniques.
The paper tackles the misalignment between standard training objectives and task-specific goals in world models by introducing RLVR-World, a framework that uses reinforcement learning with verifiable rewards to directly optimize for metrics like accuracy or perceptual quality, resulting in substantial performance gains across domains such as text games, web navigation, and robot manipulation.
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.