Research map / Label-efficient and robust learning / Recognition and long-tail
Learning Symmetric Embeddings for Equivariant World Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 an SEN with an equivariant MatMul transition instead of a non-equivariant transition | 3D Teapot world model · 3D Teapot | TH@1 yaw (%): 6.7{\color{gray}\scriptstyle \pm 6.7} -> 100{\color{gray}\scriptstyle \pm 0.0} | Helped |
| Use an SEN with an equivariant MatMul transition rather than Homeomorphic VAE embeddings | 3D Teapot world model · 3D Teapot | TH@1 yaw (%): 6.7 -> 100{\color{gray}\scriptstyle \pm 0.0} | Helped |
| Use an SEN-based equivariant model rather than a CNN | 2D Shapes world model trained on up actions · 2D Shapes (limited actions; all actions evaluation) | H@1 (10 step, %): 2.8{\color{gray}\scriptstyle \pm 0.6} -> 100{\color{gray}\scriptstyle \pm 0.0} | Helped |
| Learn a symmetric embedding with an equivariant transition rather than a non-equivariant model | 3D Blocks world model trained with limited actions · 3D Blocks (limited actions; all actions evaluation) | H@1 (10 step, %): 52.3{\color{gray}\scriptstyle \pm 14.} -> 99.9{\color{gray}\scriptstyle \pm 0.0} | Helped |
| Learn a symmetric embedding instead of imposing the inaccurate pixel-level equivariance directly | 3D Blocks world model trained with limited actions · 3D Blocks (limited actions; all actions evaluation) | H@1 (10 step, %): 83.7{\color{gray}\scriptstyle \pm 36.} -> 99.9{\color{gray}\scriptstyle \pm 0.0} | Helped |
| Use the SEN-based model instead of the non-equivariant model | Rush Hour world model · Rush Hour (10 step) | Hits@1 (%): 98.2{\color{gray}\scriptstyle \pm 1.3} -> 93.6{\color{gray}\scriptstyle \pm 3.7} | Hurt |
| Use a symmetric embedding rather than imposing inaccurate pixel-level equivariance | Rush Hour world model · Rush Hour (10 step) | Hits@1 (%): 87.9{\color{gray}\scriptstyle \pm 3.1} -> 93.6{\color{gray}\scriptstyle \pm 3.7} | Helped |
| Use the SEN-based model instead of the non-equivariant model | Reacher world model trained on positive second-joint forces · Reacher (1 step) | MRR (%): 50.6{\color{gray}\scriptstyle \pm 3.1} -> 59.4{\color{gray}\scriptstyle \pm 4.6} | Helped |
| Use the SEN-based model rather than the fully equivariant model | Reacher world model trained on positive second-joint forces · Reacher (1 step) | MRR (%): 61.8{\color{gray}\scriptstyle \pm 13.} -> 59.4{\color{gray}\scriptstyle \pm 4.6} | Mixed |
| Use the SEN-based model rather than the non-equivariant model | 3D Blocks world model with full action coverage · 3D Blocks (10 step) | Hits@1 (%): 94.3{\color{gray}\scriptstyle \pm 9.0} -> 99.9{\color{gray}\scriptstyle \pm 0.0} | No clear effect |
| Use an SEN rather than no symmetry-aware embedding | Reacher world model · Reacher (1 step) | H@10 (%): 100{\color{gray}\scriptstyle \pm 0.0} -> 100{\color{gray}\scriptstyle \pm 0.0} | No clear effect |
| Use an SEN-based model rather than no symmetry | 3D Blocks world model trained with limited actions · 3D Blocks (limited actions; all actions evaluation) | DIE (10 step, ×10^-3): 181{\color{gray}\scriptstyle \pm 79.} -> 5{\color{gray}\scriptstyle \pm 4.7} | Helped |
Closest papers by meaning
- Improving Equivariant Networks with Probabilistic Symmetry Breaking (ICLR 2025 · similarity 0.91)
- A Tale of Two Symmetries: Exploring the Loss Landscape of Equivariant Models (NeurIPS 2025 · similarity 0.90)
- Learning (Approximately) Equivariant Networks via Constrained Optimization (NeurIPS 2025 · similarity 0.89)
- On Making $SE(2)$-Invariant Networks Optimal (NeurIPS 2026 · similarity 0.89)
- Probing Equivariance and Symmetry Breaking in Convolutional Networks (NeurIPS 2025 · similarity 0.89)
- Neural Isometries: Taming Transformations for Equivariant ML (NeurIPS 2024 · similarity 0.89)
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.