Research map / Large language models / Language, safety and interpretability
GLAM: Global-Local Variation Awareness in Mamba-based World Model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GLAM, a Mamba-based world model with parallel global (GMamba) and local (LMamba) variation-aware reasoning modules, evaluated on 26 Atari 100k games.
Problem
Model-based RL world models have limited imagination quality and sample efficiency because they predict from known states directly and do not exploit state-to-state variation, or they process long sequences inefficiently.
Approach
Two parallel Mamba modules: GMamba models variation patterns over a long (16-step) history to predict future variation, while LMamba predicts rewards, terminations and visual features from variation between adjacent states in a short (4-step) window; their outputs are fused, and a phased schedule increases imagined rollout horizon during agent training.
Key results
- GLAM reaches 130.6% human mean and 61.8% human median on the 26-game Atari 100k subset, above STORM's 122.3% and 58.4%.
- GLAM scores 32680 on Kung Fu Master and 13190 on Battle Zone, versus 26182 and 13540 for STORM on those games (STORM is higher on Battle Zone).
- Ablation: replacing Transformer in STORM with Mamba (ours w/o G) lifts Pong to 14 and Boxing to 71 versus Dbl.Transformer's 4 and 68.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GLAM instead of Hieros | Atari 100k · Atari 100k (26 games) | human-normalized mean: 120% -> 130.6% | Helped |
| Use GLAM instead of STORM | Atari 100k · Atari 100k (26 games) | human-normalized median: 58.4% -> 61.8% | Helped |
| Add GMamba and LMamba to the baseline inference module (ablation) | GLAM · Pong | score: 8 -> 18 | Helped |
| Add GMamba and LMamba to the baseline inference module (ablation) | GLAM · Battle Zone | score: 8320 -> 13190 | Helped |
| Add GMamba to the LMamba-based model (ablation) | GLAM · Pong | score: 14 -> 18 | Helped |
| Add GMamba to the LMamba-based model (ablation) | GLAM · Kung Fu Master | score: 29475 -> 32680 | Helped |
| Add variation loss L_VAR (ablation) | GLAM · Pong | score: 11 -> 18 | Helped |
| Add variation loss L_VAR (ablation) | GLAM · Boxing | score: 75 -> 76 | Helped |
| Use two identical Transformer modules in parallel instead of GLAM inference (ablation) | GLAM · Pong | score: 18 -> 4 | Hurt |
| Use two identical Mamba modules in parallel instead of GLAM inference (ablation) | GLAM · Pong | score: 18 -> 3 | Hurt |
| Increase maximum imagined interaction horizon (ablation) | GLAM · Kung Fu Master | performance | Mixed |
| Vary the number of Mamba layers in GMamba and LMamba (ablation) | GLAM · Kung Fu Master | performance | Hurt |
Surprising findings
- GLAM scores 100 on Private Eye versus 7781 for STORM, and is weaker than several baselines on Asterix, Seaquest and Kangaroo.
- Stacking two identical parallel modules (Dbl.Mamba) performs worse than Dbl.Transformer on several games, so the gain comes from the specialized GMamba/LMamba design rather than parallel capacity alone.
Limitations
- Struggles to predict variation for objects that blend into the background (e.g., Crazy Climber, Alien).
- Performance declines with larger Mamba layer counts, attributed to parallel-module synchronization difficulty and error propagation.
- Evaluated only on a 26-game Atari subset; limited to variation-based reasoning in chaotic or nonlinear systems.
Experimental setup
Datasets: Atari 100k
Benchmarks: Atari 100k
Metrics: human-normalized mean score human-normalized median score per-game score
Baselines: SimPLe TWM IRIS DreamerV3 Hieros STORM
Builds on: STORM DreamerV3 Mamba IRIS
Closest papers by meaning
- RLVR-World: Training World Models with Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
- Decision Mamba: Reinforcement Learning via Hybrid Selective Sequence Modeling (NeurIPS 2024 · similarity 0.85)
- Scaling Offline Model-Based RL via Jointly-Optimized World-Action Model Pretraining (ICLR 2025 · similarity 0.85)
- Novelty Detection in Reinforcement Learning with World Models (ICML 2025 · similarity 0.85)
- Decision Mamba: A Multi-Grained State Space Model with Self-Evolution Regularization for Offline RL (NeurIPS 2024 · similarity 0.85)
- WALL-E: World Alignment by NeuroSymbolic Learning improves World Model-based LLM Agents (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.