The Verifier is the Curriculum: Execution-Gated Self-Distillation for Cross-Family Game Generation

Short summary

This paper shows that rejection-sampling self-distillation under a strict, ungameable verifier (whether a generated Godot project launches cleanly) dramatically improves cross-family game generation. A Qwen3-14B model distilled under this launch gate raised clean generation on unseen game families from 8.8% to 42.2% and achieved full best-of-K coverage. Swapping the strict verifier for a lenient one erased all gains, proving that verifier precision—not data quantity—drives improvement.

• •Strict launch-gate verifier drives cross-family game generation from 8.8% to 42.2% clean rate
• •Lenient verifier swap erases all gains, isolating verifier precision as the key factor
• •Gold-duplication control regresses below base model, confirming gains are functional not just more data

Generated with AI, which can make mistakes.