Reproduce “Zero-shot World Models via Search in Memory” on OpenResearch

About this paper

Federico Malato, Ville Hautamäki

Researchers at the University of Eastern Finland developed zero-shot world models that predict environmental dynamics by leveraging similarity search on latent representations from a pre-trained Variational Autoencoder (VAE), circumventing explicit dynamics training. This approach frequently surpasses learning-based baselines in prediction quality and demonstrates higher data efficiency, particularly for long-horizon forecasting.

Key insights

• Memory-based similarity search on latent representations can effectively approximate environmental dynamics without requiring explicit training of a dynamics model.
• Probabilistic search (Replay-KL) and temporally-aware search (Rollout) in the latent space lead to more consistent and less hallucinatory long-horizon predictions compared to simpler distance metrics.
• Zero-shot world models offer a data-efficient and computationally lightweight alternative to traditional learning-based world models, especially for tasks with adequately covered latent state spaces.

Summary by alphaXiv.