Execution-Gated Self-Distillation Boosts Game Generation by 33 Percentage Points

Jul 14, 2026

A new arXiv paper presents 'strict-launch' filtering for code generation, where only projects that launch cleanly under a headless engine are used for self-distillation. Applied to a 14B model, this approach increased clean generation on unseen game families from 8.8% to 42.2% per candidate and achieved perfect best-of-K coverage (25/25) over three rounds. The improvements are attributed to the precision of the verifier rather than simply increasing data quantity.

Why it matters: This work shows that precise, ungameable filters can significantly improve out-of-family generalization in code generation.

Full story at: arXiv AI/ML ↗