M^3 : A Modular World Model over Streams of Tokens
TLDR
Simulus is a modular token-based world model agent combining four components to achieve state-of-the-art sample efficiency in planning-free RL across three benchmarks.
Reasoning
The paper presents a clear modular approach with strong empirical results on diverse benchmarks, but the abstract lacks details on limitations and the tokenization framework's generality. The claim of synergistic gains is supported by ablation studies.
Read-first score
Read-first score 68.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 36.
Field roles
Rank sensitivity
Stability: volatile; rank range: 344.
Keyword Scores
Deep Analysis
Innovations
- Flexible tokenization framework supporting arbitrary combinations of observation and action modalities
- Intrinsic motivation for epistemic uncertainty reduction
- Prioritized world model replay
- Regression-as-classification for reward and return prediction
Methodology
Simulus is a modular token-based world model agent that integrates four components: a flexible tokenization framework, intrinsic motivation, prioritized world model replay, and regression-as-classification. It is evaluated on three diverse benchmarks (visual Atari 100K, continuous-control DMC Proprioception 500K, and symbolic Craftax-1M) against planning-free world model baselines, using sample efficiency as the primary metric.
Key Results
Simulus achieves state-of-the-art sample efficiency for planning-free world models across all three benchmarks. Ablation studies confirm that each component contributes individually and their combination yields synergistic gains.