Zero-shot World Models via Search in Memory
TLDR
Proposes a zero-shot search-based world model using similarity search and stochastic representations, achieving comparable performance to trained models with stronger long-horizon prediction.
Reasoning
The paper introduces a novel training-free world model via memory search, showing competitive results against PlaNet, especially in long-horizon dynamics. However, it only compares with one baseline and lacks real-world validation, limiting generalizability.
Read-first score
Read-first score 60.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 52.
Field roles
Rank sensitivity
Stability: volatile; rank range: 371.
Keyword Scores
Deep Analysis
Innovations
- Zero-shot world model via search in memory without training
- Use of similarity search and stochastic representations to approximate transition dynamics
- Demonstrates stronger long-horizon prediction performance compared to training-based world models
Methodology
The paper proposes a world model that leverages similarity search and stochastic representations to approximate environment dynamics without any training procedure. It compares against PlaNet, a well-established world model from the Dreamer family, evaluating on latent reconstruction quality, perceived similarity of reconstructed images, and both next-step and long-horizon dynamics prediction across visually diverse environments.
Key Results
The search-based world model achieves comparable performance to the training-based PlaNet on next-step and long-horizon prediction, and notably shows stronger performance in long-horizon prediction across a range of visually different environments.