Zero-shot World Models via Search in Memory

2025-10-17

AI Summary

This paper proposes a training-free, zero-shot world model that approximates environment transition dynamics solely via similarity search and random representation, addressing the heavy reliance of conventional trained world models (e.g., PlaNet, Dreamer) on extensive interaction data. Methodologically, it abandons explicit parametric learning; instead, it retrieves semantically similar latent states directly from a memory bank and employs random projections for efficient dynamic approximation and image reconstruction. Its key contribution is the first integration of nonparametric similarity search into world model construction—eliminating gradient-based optimization and explicit latent space learning. Experiments across diverse vision-based RL environments with substantial visual disparities demonstrate that the model achieves single-step prediction quality comparable to PlaNet, superior long-horizon forecasting performance, and significantly improved sample efficiency in online learning.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Search and Optimization: Learning to Search
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

World Models have vastly permeated the field of Reinforcement Learning. Their ability to model the transition dynamics of an environment have greatly improved sample efficiency in online RL. Among them, the most notorious example is Dreamer, a model that learns to act in a diverse set of image-based environments. In this paper, we leverage similarity search and stochastic representations to approximate a world model without a training procedure. We establish a comparison with PlaNet, a well-established world model of the Dreamer family. We evaluate the models on the quality of latent reconstruction and on the perceived similarity of the reconstructed image, on both next-step and long horizon dynamics prediction. The results of our study demonstrate that a search-based world model is comparable to a training based one in both cases. Notably, our model show stronger performance in long-horizon prediction with respect to the baseline on a range of visually different environments.

Problem

• Approximating world models without training using similarity search
• Comparing search-based models with trained ones like PlaNet
• Evaluating latent reconstruction and long-horizon prediction performance

Innovation

• Search-based world model without training procedure
• Uses similarity search and stochastic representations
• Approximates transition dynamics via memory retrieval