Efficient World Models with Context-Aware Tokenization
TLDR
Δ-IRIS uses delta encoding and context-aware tokenization for efficient transformer-based world models, achieving SOTA on Crafter with faster training.
Reasoning
Strengths: Novel delta encoding reduces sequence length, leading to significant speedup and SOTA results on Crafter. Weaknesses: Only evaluated on a single simulated benchmark (Crafter), lacking real-world or diverse environment validation.
Read-first score
Read-first score 69.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 51.
Field roles
Rank sensitivity
Stability: volatile; rank range: 107.
Keyword Scores
Deep Analysis
Innovations
- Discrete autoencoder that encodes stochastic deltas between consecutive time steps
- Autoregressive transformer that predicts future deltas by summarizing the current world state with continuous tokens
- Context-aware tokenization enabling efficient world model representation
Methodology
The paper proposes Δ-IRIS, a model-based RL agent whose world model consists of a discrete autoencoder that encodes stochastic deltas between time steps and an autoregressive transformer that predicts future deltas by summarizing the current state with continuous tokens. The agent is trained and evaluated on the Crafter benchmark across multiple frame budgets, with comparisons to prior attention-based approaches.
Key Results
Δ-IRIS achieves a new state of the art on the Crafter benchmark at multiple frame budgets while being an order of magnitude faster to train than previous attention-based world models.