{
  "aliases": [],
  "authors": [
    "Hang He",
    "Linting Wang",
    "Guanming Liu",
    "Bo Mao",
    "Jinxiu Liu",
    "Yiran Suo",
    "Maosen Zhou",
    "Zhiheng Xi",
    "Guanting Dong",
    "Zhou Shao",
    "Rui Zheng",
    "Jie Zhou",
    "Chengcheng Wan",
    "Tao Gui",
    "Qi Zhang",
    "Xipeng Qiu",
    "Xuanjing Huang"
  ],
  "details": {
    "arxiv_snapshot": {
      "abstract_hash_method": "sha256-utf8-whitespace-collapsed-v1",
      "abstract_sha256": "6add363f25f87ea1814c7b79ebf6b51f7431b3745a7d2d910960b78e2d97d13f",
      "submitted_at": "2026-10-13T09:52:02Z",
      "updated_at": "2026-10-13T09:52:02Z",
      "version": 1,
      "versioned_url": "https://arxiv.org/abs/2610.18067v1"
    },
    "license_review": {
      "document": {
        "evidence_urls": [
          "https://arxiv.org/html/2610.18067v1"
        ],
        "reason": "Document declaration only; code, data and weights permissions are not inferred."
      }
    }
  },
  "environments": [
    "unspecified"
  ],
  "evidence": [
    {
      "method": "arxiv_api",
      "retrieved_at": "2026-10-17T12:31:39Z",
      "role": "primary",
      "sha256": "c52e4e283e0dae88ef8b3ce67b7424b1d4b48e8f41015d872f4da69f98782eaa",
      "supports": [
        "name",
        "authors",
        "published_at",
        "updated_at",
        "identifiers.arxiv",
        "details.arxiv_snapshot"
      ],
      "url": "https://export.arxiv.org/api/query?id_list=2609.39924%2C2610.00346%2C2610.00583%2C2610.01939%2C2610.05099%2C2610.09127%2C2610.09484%2C2610.09633%2C2610.11251%2C2610.11922%2C2610.13098%2C2610.13166%2C2610.13361%2C2610.13623%2C2610.16117%2C2610.17953%2C2610.18067%2C2610.19366%2C2610.20424&max_results=19"
    },
    {
      "method": "editorial_primary_review",
      "retrieved_at": "2026-10-20T19:02:20Z",
      "role": "primary",
      "sha256": "f1fbdd2ecb2a54b7723372118e2b0ac818420f3a3dd280c2486c8e064cf486d6",
      "supports": [
        "name",
        "kind",
        "summary",
        "scope",
        "environments",
        "topics",
        "notes",
        "licenses.document"
      ],
      "url": "https://arxiv.org/html/2610.18067v1"
    }
  ],
  "id": "arxiv:2610.18067",
  "identifiers": {
    "arxiv": "2610.18067"
  },
  "kind": "paper",
  "last_verified_at": "2026-10-20T19:02:20Z",
  "licenses": {
    "code": null,
    "data": null,
    "document": {
      "basis": "The exact-v1 arXiv HTML infobox explicitly declares CC BY 4.0 for this paper.",
      "declared": "CC-BY-4.0",
      "retrieved_at": "2026-10-20T19:02:20Z",
      "url": "https://arxiv.org/html/2610.18067v1"
    },
    "weights": null
  },
  "links": [
    {
      "rel": "primary",
      "url": "https://arxiv.org/abs/2610.18067"
    }
  ],
  "name": "WorkRover: Scaling Long-Horizon Cowork Agent Training Via Recursive Environment Evolution",
  "notes": "App WorkEnvs expose UI observations and interaction interfaces alongside file and tool workflows; no specific operating system or pixel-control success is established. Verifier shaping depends on deterministic-check and dependency coverage. External reference scores use differing harnesses or subsets; GDPval Elo is not comparable to percentage metrics. A successful-trace analysis subset is separate from online RL outcomes. No implementation release or independent reproduction was verified.",
  "published_at": "2026-10-13",
  "relations": [],
  "review_status": "curated",
  "scope": "core",
  "summary": "Synthesizes resettable workspaces with files, application observations and skills; repairs environment incompatibilities and trains long-task agents using verifier-linked turn credit.",
  "topics": [
    "data-generation",
    "long-horizon",
    "evaluation",
    "reward-models",
    "reinforcement-learning",
    "workflow"
  ],
  "updated_at": "2026-10-13",
  "url": "https://arxiv.org/abs/2610.18067"
}