Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Building spatial world models from sparse transitional episodic memories
Zizhan He, Maxime Daigle, Pouya Bashivan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use a transformer instead of LSTM or Mamba for the world model (ablation) | ESWM · Open Arena | generalization accuracy | Improved |
| use ESWM with disjoint memories instead of TEM-T sequence model | ESWM-T · Open Arena | accuracy | Improved |
| increase model size | ESWM-T vs TEM-T · Open Arena | accuracy | Mixed |
| increase number of attention heads beyond 8 (ablation) | ESWM-T · Open Arena | accuracy | Hurt |
| increase model depth (ablation) | ESWM-T · Open Arena | accuracy | Improved |
| use denser non-minimal memory banks | ESWM · Random Wall | prediction accuracy | Improved |
| use ESWM world model with planning instead of EPN | ESWM · Random Wall navigation | success rate / optimality: 96.8% success, 99.2% optimality, +18% and +21% over EPN | Improved |
| use ESWM explorer instead of EPN | ESWM · Random Wall exploration | unique states visited: 16.8% more unique states | Improved |
| update memory bank instead of retraining after structural changes | ESWM · Random Wall with added wall | navigation success: 93 vs 72 (EPN) and 56 (baseline) | Improved |
| add contradictory observations to memory banks (noise) | ESWM · Random Wall | prediction accuracy: dropped by 1% and 10% | Hurt |
| train on episodic transitions to build latent map (trained vs untrained) | ESWM · sRSA | sRSA: 0.5 -> 0.78 | Improved |
| hand-engineered memory lookup instead of learnable attention (ablation) | hand-engineered lookup · Open Arena | accuracy | Hurt |
Closest papers by meaning
- Time Makes Space: Emergence of Place Fields in Networks Encoding Temporally Continuous Sensory Experiences (NeurIPS 2024 · similarity 0.91)
- Does Spatial Cognition Emerge in Frontier Models? (ICLR 2025 · similarity 0.88)
- Spatially-Aware Transformers for Embodied Agents (ICLR 2024 · similarity 0.88)
- Global Distortions from Local Rewards: Neural Coding Strategies in Path-Integrating Neural Systems (NeurIPS 2024 · similarity 0.88)
- Video World Models with Long-term Spatial Memory (NeurIPS 2025 · similarity 0.88)
- A simple model of co-emergence of grid and place fields (NeurIPS 2026 · similarity 0.88)
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.
