Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Grounded Language Learning Fast and Slow
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 |
|---|---|---|---|
| Add observation reconstruction loss (ablation) | DCEM mem=1024 · Unity fast-mapping, 3 objects, |G|=30 | accuracy: 0.33 (0.02) -> 0.98 (0.01) | Helped |
| Add observation reconstruction loss (ablation) | TransformerXL mem=1024 · Unity fast-mapping, 3 objects, |G|=30 | accuracy: 0.32 (0.02) -> 0.98 (0.01) | Helped |
| Add observation reconstruction loss (ablation) | DNC mem=1024 · Unity fast-mapping, 3 objects, |G|=30 | accuracy: 0.34 (0.01) -> 0.64 (0.27) | Helped |
| Add observation reconstruction loss (ablation) | LSTM · Unity fast-mapping, 3 objects, |G|=30 | accuracy: 0.33 (0.05) -> 0.61 (0.27) | Helped |
| Reduce memory window/buffer from 1024 to 100 (ablation) | DCEM + R · Unity fast-mapping, 3 objects, |G|=30 | accuracy: 0.98 (0.01) -> 0.98 (0.01) | No clear effect |
| Reduce memory window/buffer from 1024 to 100 (ablation) | TransformerXL + R · Unity fast-mapping, 3 objects, |G|=30 | accuracy: 0.98 (0.01) -> 0.73 (0.35) | Hurt |
| Use DCEM rather than TransformerXL with a 100-step memory | Unity fast-mapping, 3 objects, |G|=30 · Unity fast-mapping, 3 objects, |G|=30 | accuracy: 0.73 (0.35) -> 0.98 (0.01) | Helped |
| Meta-train with more objects per episode | DCEM + reconstruction · Fast-mapping episodes with more objects than training | accuracy | Helped |
| Train with category exemplars paired across discovery and instruction | DCEM + reconstruction · ShapeNet fast category extension | accuracy: around 55% -> 88% | Helped |
| Read multiple memories per query (k > 1) (ablation) | DCEM + reconstruction · Novel-object generalization | accuracy | Helped |
| Read multiple memories per query (k > 1) (ablation) | DCEM + reconstruction · ShapeNet category extension | accuracy | Helped |
| Use DCEM memory keys, values, or both for NGU novelty reward | DCEM agent without shaping reward · Unity fast-mapping, 3 objects | accuracy | Helped |
Closest papers by meaning
- Interactive Grounded Language Acquisition and Generalization in a 2D World (ICLR 2018 · similarity 0.86)
- Learning to Model the World With Language (ICML 2024 · similarity 0.85)
- Iterated Learning Improves Compositionality in Large Vision-Language Models (CVPR 2024 · similarity 0.85)
- The Emergence of Abstract Thought in Large Language Models Beyond Any Language (NeurIPS 2025 · similarity 0.84)
- Words That Make Language Models Perceive (NeurIPS 2026 · similarity 0.84)
- Ground-Compose-Reinforce: Grounding Language in Agentic Behaviours using Limited Data (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.