{
  "abstract": "Reliable agent evaluation is complicated by automatic harness optimization, which repeatedly uses a released benchmark $B_{\\mathrm{rel}}$ to guide a Proposer that edits prompts, memory, retrieval, tools, and control code around a fixed foundation model. Task holdout is commonly used to guard against harness overfitting. It varies semantic tasks but leaves the benchmark protocol fixed, so a bad genius Proposer can produce a cheating harness whose improvement over the initial harness on $B_{\\mathrm{rel}}$ depends on a benchmark-wide shortcut. We introduce Counterfactual Harness Search and Evolution (CHASE), which casts harness evolution as constraint generation over valid counterfactual benchmarks. After each Proposer update, a Challenger searches for an executable protocol transformation with large gain destruction. A validity firewall checks that task semantics are preserved, while a held-out confirmation set determines whether the counterfactual enters a finite archive. We formalize an ideal shortcut-neutralized benchmark $B_0$ and establish theoretical guarantees linking finite counterfactual archives to $B_0$ and characterizing sequential Challenger search. We evaluate CHASE on Syn-Ledger and OfficeQA, where CHASE retains strong released-benchmark gains while substantially reducing gain destruction under valid protocol transformations.",
  "abstract_shown": true,
  "affiliations": [],
  "article_type": null,
  "arxiv": "2609.18366",
  "author_total": 6,
  "authors": [
    {
      "name": "Guojun Zhu"
    },
    {
      "name": "Xunheng Huang"
    },
    {
      "name": "Peng Yin"
    },
    {
      "name": "Jiahui Xie"
    },
    {
      "name": "Sanguo Zhang"
    },
    {
      "name": "Doudou Zhou"
    }
  ],
  "bibtex": "https://research.draconitic.com/paper/25qc4a0p160wn.bib",
  "cited_by": 0,
  "doi": "10.48550/arxiv.2609.18366",
  "fulltext_url": null,
  "id": "25qc4a0p160wn",
  "issns": [],
  "landing_url": "https://arxiv.org/abs/2609.18366",
  "language": null,
  "licence": "LicenseRef-arXiv-nonexclusive-1.0",
  "licence_label": "arXiv licence",
  "licence_source": "article",
  "licence_url": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/",
  "metadata_licence": "CC0-1.0",
  "native_subjects": [
    [
      "arxiv",
      "cs.AI"
    ],
    [
      "arxiv",
      "cs.LG"
    ],
    [
      "arxiv",
      "stat.ML"
    ],
    [
      "openalex",
      "1702|Evolutionary Algorithms and Applications"
    ]
  ],
  "openalex": "W7213492889",
  "pdf_url": "https://arxiv.org/pdf/2609.18366",
  "pmcid": null,
  "pmid": null,
  "published": "2026-09-24",
  "publisher": null,
  "ris": "https://research.draconitic.com/paper/25qc4a0p160wn.ris",
  "sources": [
    {
      "first_seen": "2026-10-30T12:24:21Z",
      "native_id": "2609.18366",
      "source": "arxiv",
      "url": "https://arxiv.org/abs/2609.18366"
    }
  ],
  "subjects": [
    {
      "field": "mathematics",
      "origin": "arxiv:stat.ML",
      "subfield": "statistics"
    }
  ],
  "title": "Bad Genius: Counterfactual-Guided Harness Evolution Beyond Task-Specific Shortcuts",
  "title_text": "Bad Genius: Counterfactual-Guided Harness Evolution Beyond Task-Specific Shortcuts",
  "type": "preprint",
  "updated_at": "2026-10-30T12:24:21Z",
  "url": "https://research.draconitic.com/paper/25qc4a0p160wn",
  "venue": "arXiv",
  "versions": []
}