[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-85125-en":3,"doc-seo-85125-105":30,"detail-sidebar-cat-0-en-105":91},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":21,"is_downloadable":21,"audit_status":21,"page_count":22,"language":23,"language_code":24,"site_id":25,"html_lang":24,"table_of_contents":26,"faqs":27,"seo_title":13,"seo_description":14,"update_tm":28,"read_time":29},85125,2336464648746,"Skyler","https://ap-avatar.wpscdn.com/davatar_276721f389ce27ea32af1340a28f341c",8,"Research & Report","The Verifier is the Curriculum: Execution-Gated Self-Distillation for Cross-Family Game Generation","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.","arXiv :2607 .09709v 1 [ cs .AI] 23 Jun 2026  \nThe Verifier is the Curriculum:  \nExecution-Gated Self-Distillation for Cross-Family  \nGame Generation  \nChenyu Zhou 1 Qiliang Jiang2 Shuning Wu3 Xu Zhou3 ∗  \n1 School of Engineering, Institute of Science Tokyo, Japan  \n2 College of Control Science and Engineering, Zhejiang University, China  \n3 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)  \nAbstract  \nPost-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,  \njudge-free, ungameable filter—whether a generated project launches cleanly under a headless  \nengine (strict-launch) . Under this gate, rejection-sampling self-distillation compounds out  \nof-family generalization. On GameCraft-Bench (mapping a natural-language brief to a complete  \nGodot project), a 14B model (Qwen3-14B+LoRA) distilled under strict-launch raises clean  \ngeneration on four unseen game families from 8 .8% to 42.2% per-candidate and best-of-K  \ncoverage from 18/25 to 25/25 (the gold ceiling) over three rounds, each a significant gain  \n(p=0 .0019, p\u003C10−4 , p\u003C10−4) . The gain is not from merely adding data: an exactly-matched  \ngold-duplication control regresses below the base model (5 .6% vs. 8.8%, p = 0 .019), while a  \ncount-matched decomposition splits the round-1-to-2 jump into comparable quality (+8 .8pp)  \nand quantity (+8 .5pp) channels. Most directly, rerunning the loop with only the filter swapped—  \nthe lenient BUILD check, which passes 99 .9% of generations, in place of the launch gate—erases  \nthe gain entirely (back to base, p=10−3 vs. the launch-gated round), isolating verifier precision  \nrather than the optimizer. A second ungameable signal, headless execution grounding, rises  \nmonotonically across rounds and yields far more grounded candidates than gold-duplication at  \na matched budget (16 vs. 5), confirming the gains are functional, not launch-but-empty. Game  \ngeneration is a verifiable testbed for one lesson: the verifier is the curriculum—what it certifies  \nis what the model learns.  \n1 Introduction  \nPost-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  \n∗ Corresponding author: zhouxu [nus@u.nus.edu](nus@u.nus.edu)  \nagent 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.  \nThis 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.  \nThe task we","cbCaivEYObWGhr9a","https://ap.wps.com/l/cbCaivEYObWGhr9a","pdf",463994,2,1,15,"English","en",105,"# Abstract\n# Introduction\n## Learned judges and gaming failure modes\n## Execution-gated verifier and strict-launch\n## GameCraft task and verifiable success criteria\n## Key findings: compounding gains from rejection-sampling self-distillation","[{\"question\":\"What problem does the paper address with learned judges in code generation post-training?\",\"answer\":\"Learned judges can be gamed by optimizing for proxy surface features that raise scores without improving the underlying artifact.\"},{\"question\":\"What is the paper’s proposed verification signal for self-distillation?\",\"answer\":\"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).\"},{\"question\":\"How does the strict-launch verifier affect performance on unseen game families?\",\"answer\":\"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.\"}]",1784201240,38,{"code":4,"msg":31,"data":32},"ok",{"site_id":25,"language":24,"slug":33,"title":13,"keywords":34,"description":14,"schema_data":35,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":28},"the-verifier-is-the-curriculum-execution-gated-self-distillation-for-cross-family-game-generation","",{"@graph":36,"@context":85},[37,53,68],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,47,50],{"item":41,"name":42,"@type":43,"position":21},"https://docshare.wps.com","Home","ListItem",{"item":45,"name":46,"@type":43,"position":20},"https://docshare.wps.com/document/","Document",{"item":48,"name":12,"@type":43,"position":49},"https://docshare.wps.com/document/research-report/",3,{"item":51,"name":13,"@type":43,"position":52},"https://docshare.wps.com/document/the-verifier-is-the-curriculum-execution-gated-self-distillation-for-cross-family-game-generation/85125/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":24,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":41,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-24","2026-07-16",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"What problem does the paper address with learned judges in code generation post-training?","Question",{"text":75,"@type":76},"Learned judges can be gamed by optimizing for proxy surface features that raise scores without improving the underlying artifact.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"What is the paper’s proposed verification signal for self-distillation?",{"text":80,"@type":76},"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).",{"name":82,"@type":73,"acceptedAnswer":83},"How does the strict-launch verifier affect performance on unseen game families?",{"text":84,"@type":76},"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.","https://schema.org",{"og:url":51,"og:type":87,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":89,"canonical":51},"index,follow",{"doc_id":7,"site_id":25},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,135],{"id":21,"doc_module":4,"doc_module_name":46,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":20,"doc_module":4,"doc_module_name":46,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":46,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":46,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":46,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":46,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":46,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":46,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":46,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":46,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]