AI learns game generation via execution-gated self-distillation

Researchers have developed a novel self-distillation technique for training AI models to generate functional game projects from natural language descriptions. This method, termed "execution-gated self-distillation," uses a strict launch check as a filter, ensuring generated projects can run without errors. When applied to the GameCraft-Bench dataset using a Qwen3-14B model, this approach significantly improved the generation of complete Godot projects across unseen game families, raising success rates from 8.8% to 42.2% per candidate. AI

IMPACT This method could improve AI's ability to generate functional and complex code, particularly in creative domains like game development.

RANK_REASON The cluster contains a research paper detailing a new AI training methodology. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• alphaXiv
• CatalyzeX
• DagsHub
• GameCraft-Bench
• Godot
• Gotit.pub
• Hugging Face
• Lora
• Qwen3 14B
• ScienceCast

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →