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

验证器即课程：用于跨系列游戏生成的执行门控自蒸馏

Abstract: Post-training a code generator against a learned judge can optimize proxy features that raise the score without improving the artifact. We study the opposite signal: a deterministic, judge-free, ungameable filter — whether a generated project launches cleanly under a headless engine (strict-launch).

摘要： 针对已学习的判别器（learned judge）对代码生成器进行后训练，可能会优化那些仅能提高分数却无法改善实际产物的代理特征。我们研究了一种相反的信号：一种确定性的、无需判别器的、不可作弊的过滤器——即生成的项目是否能在无头引擎（headless engine）下顺利启动（严格启动，strict-launch）。

Under this gate, rejection-sampling self-distillation compounds out-of-family generalization. On GameCraft-Bench (mapping a natural-language brief to a complete Godot project), a 14B model (Qwen3-14B+LoRA) distilled under strict-launch raises clean generation on four unseen game families from 8.8% to 42.2% per-candidate and best-of-K coverage from 18/25 to 25/25 (the gold ceiling) over three rounds, each a significant gain (p=0.0019, p<1e-4, p<1e-4).

在这种门控机制下，拒绝采样自蒸馏（rejection-sampling self-distillation）增强了跨系列（out-of-family）的泛化能力。在 GameCraft-Bench（将自然语言简报映射为完整的 Godot 项目）测试中，一个经过严格启动门控蒸馏的 14B 模型（Qwen3-14B+LoRA），在三个轮次后，将四个未见过的游戏系列中的干净生成率从每个候选 8.8% 提升至 42.2%，并将 Best-of-K 覆盖率从 18/25 提升至 25/25（达到黄金标准上限），每一轮均取得了显著增长（p=0.0019, p<1e-4, p<1e-4）。

The gain is not from merely adding data: an exactly-matched gold-duplication control regresses below the base model (5.6% vs. 8.8%, p=0.019), while a count-matched decomposition splits the round-1-to-2 jump into comparable quality (+8.8pp) and quantity (+8.5pp) channels.

这种提升并非仅仅源于数据增加：一项完全匹配的黄金数据复制对照实验显示其性能反而低于基础模型（5.6% 对比 8.8%，p=0.019）；而一项数量匹配的分解实验则将第一轮到第二轮的跳跃拆解为可比的质量提升（+8.8pp）和数量提升（+8.5pp）两个维度。

Most directly, rerunning the loop with only the filter swapped — the lenient BUILD check, which passes 99.9% of generations, in place of the launch gate — erases the gain entirely (back to base, p=1e-3 vs. the launch-gated round), isolating verifier precision rather than the optimizer.

最直接的证据是，如果仅替换过滤器——使用通过率高达 99.9% 的宽松构建检查（BUILD check）来代替启动门控——之前的增益将完全消失（回归到基础水平，p=1e-3，对比启动门控轮次），这证明了验证器的精度而非优化器本身才是关键。

A second ungameable signal, headless execution grounding, rises monotonically across rounds and yields far more grounded candidates than gold-duplication at a matched budget (16 vs. 5), confirming the gains are functional, not launch-but-empty. Game generation is a verifiable testbed for one lesson: the verifier is the curriculum — what it certifies is what the model learns.

第二个不可作弊的信号——无头执行接地（headless execution grounding）——在各轮次中单调上升，并且在相同预算下产生的接地候选者远多于黄金数据复制（16 对比 5），这证实了这些增益是功能性的，而非仅仅是“能启动但内容为空”。游戏生成为我们提供了一个可验证的测试平台，揭示了一个道理：验证器即课程——它所认证的内容，正是模型所学习的内容。