---
格式版本: 2
标题: "Procedural Content Metageneration via Program Search and Continual Abstraction Discovery"
原文链接: "https://arxiv.org/abs/2608.17947"
发布日期: "2026-08-18"
发布时间校准状态: "found"
发布时间需复核: "否"
发布时间来源: "rule:local:strict_original_body"
发布时间证据: "**\\[v1\\]** Tue, 18 Aug 2026 16:03:22 UTC (2,585 KB)"
发布时间校准原因: "规则确认唯一严格发布时间，来源 local:strict_original_body"
发布时间校准置信度: "high"
发布时间候选数量: 18
发布时间严格候选数量: 6
发布时间原页读取状态: "source template page reused from URL open"
发布时间未找到原因: ""
发布时间校准时间: "2026-08-20T15:54:01+08:00"
发布时间仲裁状态: "skipped"
发布时间仲裁尝试次数: 0
发布时间仲裁耗时毫秒: 0
发现时间: "2026-08-20T15:48:16+08:00"
入库时间: "2026-08-20T07:54:01.211Z"
来源平台: "arXiv 学术论文搜索"
搜索渠道: "source_template"
搜索词: "https://arxiv.org/search/?query=model%20fit&searchtype=all"
匹配关键词:
  - "model fit"
  - "AI"
相关厂家:
  []
相关专家:
  []
内容类型: "网页"
抓取工具: "Free Fetch + Defuddle"
清洗工具: "Defuddle Markdown + Defuddle/Readability 正文提取"
原始附件:
  []
AI优质: "否"
AI打分: 3
AI分档: "非优质"
AI质检状态: "不通过"
AI打分理由: "论文主题为程序内容元生成，与超节点、AI Rack、机柜级AI基础设施完全无关，无相关技术或商业信息。"
AI质检模型: "ali-deepseek-v4-flash"
AI质检时间: "2026-08-20T16:00:15+08:00"
AI主题相关性: 0
AI来源权威性: 2
AI新颖性: 0
AI技术细节: 0
AI商业部署信号: 0
AI完整性: 1
AI摘要: "研究提出持续抽象发现（CAD），让大语言模型通过变异与交叉搜索完整Python程序，直接进化《Sokoban》《Zelda》等四款游戏的程序化内容生成器；"
AI摘要模型: "ali-deepseek-v4-flash"
AI摘要时间: "2026-09-07T03:20:49.024Z"
采集批次: "2026年8月20日14点19分32秒"
采集批次ID: "20260820-141932-079"
去重键: "https://arxiv.org/abs/2608.17947"
---

## Computer Science > Artificial Intelligence

## Title:Procedural Content Metageneration via Program Search and Continual Abstraction Discovery

Authors:[Matthew Siper](https://arxiv.org/search/cs?searchtype=author&query=Siper,+M), [Ahmed Khalifa](https://arxiv.org/search/cs?searchtype=author&query=Khalifa,+A), [Julian Togelius](https://arxiv.org/search/cs?searchtype=author&query=Togelius,+J)

[View PDF](https://arxiv.org/pdf/2608.17947) [HTML (experimental)](https://arxiv.org/html/2608.17947v1)

> Abstract:Large language models can generate executable programs, which makes it possible to search directly over procedural content generators rather than individual levels. We study this approach in Sokoban, Zelda, Dangerous Dave, and Lode Runner. Each run evolves complete Python generators through language-model mutation and crossover. We introduce Continual Abstraction Discovery, or CAD, which extracts reusable primitives from high-fitness programs into a run-specific helper module. A 2x2 experiment crosses CAD with access to a fixed hand-written domain API. The completed data set contains 160 complete runs, with at least ten 50-generation runs in every cell. CAD raises mean final best fitness in all eight domain and API comparisons. Across all CAD runs, learned libraries are adopted by most later programs and repeatedly rediscover validation, reachability, and structural utilities. These results support that discovering reusable primitives improves evolutionary program search for content generators.

| Comments: |  |
| --- | --- |
| Subjects: | Artificial Intelligence (cs.AI); Machine Learning (cs.LG); Neural and Evolutionary Computing (cs.NE) |
| Cite as: | [arXiv:2608.17947](https://arxiv.org/abs/2608.17947) \[cs.AI\] |
|  | (or [arXiv:2608.17947v1](https://arxiv.org/abs/2608.17947v1) \[cs.AI\] for this version) |
|  | [https://doi.org/10.48550/arXiv.2608.17947](https://doi.org/10.48550/arXiv.2608.17947) |

## Submission history

From: Julian Togelius \[[view email](https://arxiv.org/show-email/7ebe6745/2608.17947)\]  
**\[v1\]** Tue, 18 Aug 2026 16:03:22 UTC (2,585 KB)

[Which authors of this paper are endorsers?](https://arxiv.org/auth/show-endorsers/2608.17947) | Disable MathJax ([What is MathJax?](https://info.arxiv.org/help/mathjax.html))