Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Generating Interactive Worlds with Text
Angela Fan, Jack Urbanek, Pratik Ringshia, Emily Dinan, Emma Qian, Siddharth Karamcheti, Shrimai Prabhumoye, Douwe Kiela, Tim Rocktaschel, Arthur Szlam, Jason Weston
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 Starspace instead of Information Retrieval for location prediction | LIGHT location prediction · LIGHT test set | Hits at 1: 30.0 -> 45.5 | Improved |
| Provide location descriptions in addition to names (ablation) | Starspace location prediction · LIGHT test set | Hits at 1: 44.5 -> 45.5 | Improved |
| Use Starspace instead of Information Retrieval for object prediction | LIGHT object prediction · LIGHT test set | Hits at 1: 21.9 -> 47.3 | Improved |
| Provide location descriptions in addition to names (ablation) | Starspace object prediction · LIGHT test set | Hits at 1: 13.3 -> 47.3 | Improved |
| Use BERT Cross-Encoder instead of Starspace | LIGHT character prediction · LIGHT test set | Hits at 1: 35.7 -> 36.1 | Improved |
| Pretrain the generative Transformer on Reddit comments (ablation) | LIGHT location generation · LIGHT test set | F1: 23.3 -> 26.3 | Improved |
| Pretrain the generative Transformer on Reddit comments (ablation) | LIGHT character generation · LIGHT test set | F1: 28.4 -> 29.3 | Improved |
| Pretrain the generative Transformer on Reddit comments (ablation) | LIGHT object generation · LIGHT test set | F1: 32.5 -> 38.8 | Improved |
| Pretrain the generative Transformer on Reddit comments (ablation) | LIGHT affordance generation · LIGHT test set | F1: 57.0 -> 62.6 | Improved |
| Use Starspace-predicted location paths instead of data-created paths | LIGHT world construction · Human pairwise evaluation | preference: over 60% | Improved |
| Add model suggestions to the world-building interface | Interactive world creation · Nine-location game environment design | time spent: 10-20 minutes -> 10 minutes or less | Improved |
Closest papers by meaning
- Moonshine: Distilling Game Content Generators into Steerable Generative Models (AAAI 2025 · similarity 0.87)
- GAVEL: Generating Games via Evolution and Language Models (NeurIPS 2024 · similarity 0.84)
- Learning to Model the World With Language (ICML 2024 · similarity 0.83)
- TALES: Text-Adventure Learning Environment Suite (NeurIPS 2026 · similarity 0.83)
- GameGen-X: Interactive Open-world Game Video Generation (ICLR 2025 · similarity 0.83)
- PoE-World: Compositional World Modeling with Products of Programmatic Experts (NeurIPS 2025 · similarity 0.83)
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.
