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

Q: What problem does the paper address with learned judges in code generation post-training?

Learned judges can be gamed by optimizing for proxy surface features that raise scores without improving the underlying artifact.

Q: What is the paper’s proposed verification signal for self-distillation?

A deterministic, judge-free execution check: a generated project is accepted only if it launches cleanly under a headless engine with no parse, load, or runtime errors (strict-launch).

Q: How does the strict-launch verifier affect performance on unseen game families?

Rejection-sampling self-distillation under strict-launch compounds gains, raising clean generation from 8.8% to 42.2% per candidate and improving best-of-K coverage up to the 25/25 gold ceiling across three rounds.

Jul 16, 2026

Upload by Skyler

Post-training a code generator against a learned judge can improve scores via proxy artifacts without producing better programs. This work studies the reverse: a deterministic, judge-free, ungameable verifier that accepts only projects that launch cleanly under a fixed headless engine (strict-launch). With GameCraft-Bench, a distilled 14B model increases clean generation on unseen game families from 8.8% to 42.2% per candidate and reaches full best-of-K coverage over three rounds.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages15

Reading Time38 min

Abstract

Introduction

Learned judges and gaming failure modes

Execution-gated verifier and strict-launch

GameCraft task and verifiable success criteria

Key findings: compounding gains from rejection-sampling self-distillation

arXiv :2607 .09709v 1 [ cs .AI] 23 Jun 2026 The Verifier is the Curriculum: Execution-Gated Self-Distillation for Cross-Family Game Generation Chenyu Zhou 1 Qiliang Jiang2 Shuning Wu3 Xu Zhou3 ∗ 1 School of Engineering, Institute of Science Tokyo, Japan 2 College of Control Science and Engineering, Zhejiang University, China 3 Department of Electrical and Computer Engineering, National University of Singapore, Singapore [zhou.c.76d6@m.isct.ac.jp](zhou.c.76d6@m.isct.ac.jp) [jiangqiliang@zju.edu.cn](jiangqiliang@zju.edu.cn) [shuningwu@u.nus.edu](shuningwu@u.nus.edu) zhouxu [nus@u.nus.edu](nus@u.nus.edu) 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) . 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<10−4 , p<10−4) . 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. 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=10−3 vs. the launch-gated round), isolating verifier precision rather than the optimizer. 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. 1 Introduction Post-training a code-generation model requires a signal that says whether a candidate is good. The dominant choice for open-ended generation tasks is a learned judge—an LLM or a multimodal model prompted to score the output. This is convenient and correlates with human preference, but it has a structural failure mode: a learned judge maps an output to a score through whatever surface features drive its prediction, and if any of those features is cheap to manipulate without improving the artifact, a model optimized against the judge will find it. A companion study establishes separately (§3) that GameCraft-Bench’s official judge has exactly such a feature: an ∗ Corresponding author: zhouxu [nus@u.nus.edu](nus@u.nus.edu) agent lifts its art score by swapping solid-color placeholders for real assets while the game’s code is frozen. That learned judges and automatic benchmarks can be gamed is by now well documented [16, 22 , 23]; the constructive question we ask is what post-training does when the signal instead cannot be gamed. This paper takes the opposite stance on which signal to train against. Rather than chase abetter learned judge, we make the training filter a deterministic, judge-free, ungameable execution check and put it at the gate of an iterative self-distillation loop. A candidate passes only if the materialized project launches cleanly under a headless engine—it returns exit code 0 with no parse, load, or runtime error (strict-launch) . This signal exposes no scalar score to manipulate: it is a property of the artifact under a fixed engine, not an opinion. The task we

FAQs

What problem does the paper address with learned judges in code generation post-training?

Learned judges can be gamed by optimizing for proxy surface features that raise scores without improving the underlying artifact.

What is the paper’s proposed verification signal for self-distillation?

A deterministic, judge-free execution check: a generated project is accepted only if it launches cleanly under a headless engine with no parse, load, or runtime errors (strict-launch).

How does the strict-launch verifier affect performance on unseen game families?

Rejection-sampling self-distillation under strict-launch compounds gains, raising clean generation from 8.8% to 42.2% per candidate and improving best-of-K coverage up to the 25/25 gold ceiling across three rounds.

Jul 16, 2026

Upload by Skyler