Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Dreaming in Code for Curriculum Learning in Open-Ended Worlds
Konstantinos Mitsides, Maxence Faldor, Antoine Cully
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 |
|---|---|---|---|
| add FM-generated code-level curriculum (DiCode) to PPO-GTrXL | PPO-GTrXL on Craftax · Craftax held-out test | final mean return: 41.59 -> 48.55 | Improved |
| add learnability-prioritized parent selection instead of random parent sampling (ablation) | DiCode · Craftax held-out test | final mean return: 46.44 -> 48.55 | Improved |
| add learnability-prioritized parent selection (late-game combat) (ablation) | DiCode · Craftax Gnome Warrior / Archer success | success rate: Gnome Warrior: 0.6% -> 14.7%, Gnome Archer: 0.1% -> 11.2% | Improved |
| add environment reshaping (initial states and dynamics) beyond goal selection (ablation) | DiCode · Craftax held-out test | final mean return: 42.82 -> 48.55 | Improved |
| use larger generator FM (Qwen 80B instead of 30B) (ablation) | DiCode · Craftax held-out test | final mean return: 44.38 -> 45.88 | Improved |
| use Qwen3-235B instead of Qwen 80B as generator (ablation) | DiCode · Craftax held-out test | final mean return: 45.88 -> 48.55 | Improved |
| add closed-loop grounding (parent description and performance profiles in FM context) (ablation) | DiCode · Craftax held-out test | final mean return: 41.12 -> 48.55 | Improved |
| open-loop generation versus PPO-GTrXL baseline (ablation) | DiCode-OL · Craftax held-out test | final mean return: 41.59 -> 41.12 | No clear effect |
| DiCode versus best baseline on Make Iron Armour | PPO-GTrXL baselines · Craftax Make Iron Armour | success rate: 15% -> 47% | Improved |
| DiCode versus strongest baseline on entering Gnomish Mines | PPO-GTrXL baselines · Craftax Floor 2 entry | fraction of episodes: 9% -> 32% | Improved |
| DiCode versus baselines on late-game combat | PPO-GTrXL baselines · Craftax Gnome Warrior / Archer | success rate: 0% -> 15% (Gnome Warrior), 0% -> 11% (Gnome Archer) | Improved |
| DiCode versus baseline on Make Diamond Sword | PPO-GTrXL baselines · Craftax Make Diamond Sword | success rate: 3% -> 6% | Improved |
Closest papers by meaning
- OMNI-EPIC: Open-endedness via Models of human Notions of Interestingness with Environments Programmed in Code (ICLR 2025 · similarity 0.87)
- AgentOdyssey: Open-Ended Long-Horizon Text Game Generation for Test-Time Continual Learning Agents (NeurIPS 2026 · similarity 0.85)
- MindForge: Empowering Embodied Agents with Theory of Mind for Lifelong Cultural Learning (NeurIPS 2025 · similarity 0.85)
- Imagined Autocurricula (NeurIPS 2025 · similarity 0.85)
- Efficient Skill Grounding via Code Refactoring with Small Language Models (ICML 2026 · similarity 0.85)
- SimWorld Studio: Automatic Environment Generation with Evolving Coding Agent for Embodied Agent Learning (NeurIPS 2026 · similarity 0.85)
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.
