{
  "schema_version": 2,
  "id": "2604.23099",
  "source_urls": [
    "https://arxiv.org/html/2604.23099v1"
  ],
  "source_kind": "official_arxiv_html_or_abs",
  "capture_window_utc": {
    "start": "2026-10-20T18:48:00Z",
    "end": "2026-10-20T18:56:04Z"
  },
  "original_capture_sha256": "395e92d90fcc6b0f0fed11f404f21ac7a76330544b5a6ff09485900a560b94c7",
  "original_capture_hash_scope": "complete original UTF-8 browser excerpt receipt archived only in local private temporary storage; not the original HTTP response body",
  "public_receipt_kind": "source_metadata_locations_and_chinese_paraphrase_without_verbatim_article_or_internal_browser_references",
  "source_locations": {
    "read_sections": [
      "§2 Algorithm2",
      "§3.1数据/模型/feature",
      "Table5"
    ],
    "browser_line_ranges": "L274-L336;L452-L462"
  },
  "full_article_read": false,
  "method_paraphrase_zh": "ProEval预训练GP surrogate，Bayesian quadrature估计总体性能与superlevel set failure采样；TPF/RPF/SF迁移；topic UCB1＋LLM failure pattern转置。",
  "experiment_paraphrase_zh": "16LLM/VLM×reasoning/knowledge/safety/text/image多benchmark；failure discovery100迭代；GP使用历史外部模型score，目标test gold不能给evaluated agent。"
}