Beyond the Next Step: Variable-Length Latent World Models for Long-Horizon Planning

The paper introduces Variable-length Latent World Models (VLWMs), which enhance long-horizon planning by predicting future latent states based on variable-length action sequences, rather than relying solely on one-step predictions. This approach directly models temporally extended dynamics and incorporates a curriculum training strategy that stabilizes optimization for both short and long-range predictions. Experimental results demonstrate a 13% average improvement over the state-of-the-art LeWM on long-horizon control tasks, indicating VLWMs' potential to significantly enhance predictive capabilities in latent world models for practitioners.