{
  "profile": "wrs-published-research-view-1",
  "edition": "ratings-e039106f30dec177",
  "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS",
  "entry": {
    "id": "wen-et-al-claude-weak-to-strong-research-2026",
    "title": "Wen and colleagues with Claude: automated weak-to-strong supervision research",
    "summary": "Ran parallel research agents on weak-to-strong supervision, obtaining and testing combinations of representation-based confidence filtering, evolutionary adapter refinement and posterior-label noise modeling, with selected method transfer to new math and coding datasets. Reported result; not independently verified by WRS. Ranked using a provisional grade and reference estimate for the exact claimed result.",
    "boundary": "Repeated access makes the principal evaluation set effectively validation, and agents found shortcuts and evaluation exploits. Independent-domain transfer varies, and a production-scale posterior-label trial showed only a noise-floor improvement. Recovery on selected binary tasks is not a general solution to weak-to-strong alignment or superiority over unrestricted human research.",
    "area": "AI-assisted learning-method discovery",
    "topic": "Computing & AI",
    "year": 2026,
    "lane": "AI",
    "result": "ALGORITHM",
    "package_id": "wen-et-al-claude-weak-to-strong-research-2026",
    "target": {
      "record_id": "urn:uuid:e7f3afda-8145-45ba-a732-8f46ddbfc04c",
      "version": 1,
      "content_hash": "sha256:cf9f8ec666ce173932df1d6f5bb7158810ba9fbcf50c0bcc2b6fca0e3dffa70f"
    },
    "humans": [
      "Jiaxin Wen",
      "Liang Qiu",
      "Joe Benton",
      "Jan Hendrik Kirchner",
      "Jan Leike"
    ],
    "human_identities": [
      {
        "author_position": 0,
        "source_name": "Jiaxin Wen",
        "person_id": "jiaxin-wen-anthropic",
        "display_name": "Jiaxin Wen",
        "status": "SOURCE_LINKED_IDENTITY",
        "sources": [
          {
            "title": "Automated Researchers Can Mitigate Well-Characterized Alignment Failures",
            "url": "https://alignment.anthropic.com/2026/automated-alignment-researchers/"
          },
          {
            "title": "Automated Weak-to-Strong Researcher",
            "url": "https://alignment.anthropic.com/2026/automated-w2s-researcher/"
          }
        ]
      },
      {
        "author_position": 3,
        "source_name": "Jan Hendrik Kirchner",
        "person_id": "jan-hendrik-kirchner",
        "display_name": "Jan Hendrik Kirchner",
        "status": "SOURCE_LINKED_IDENTITY",
        "sources": [
          {
            "title": "Automated Researchers Can Mitigate Well-Characterized Alignment Failures",
            "url": "https://alignment.anthropic.com/2026/automated-alignment-researchers/"
          },
          {
            "title": "Automated Weak-to-Strong Researcher",
            "url": "https://alignment.anthropic.com/2026/automated-w2s-researcher/"
          }
        ]
      }
    ],
    "authors_complete": true,
    "human_role_note": "Named source authors and agent execution are recorded separately; individual effort shares are unknown.",
    "agents": [
      {
        "provider": "Anthropic",
        "product": "Claude",
        "model": "Claude Opus 4.6",
        "role": "Source-reported autonomous research agents proposing, implementing and testing weak-to-strong learning methods in human-designed sandboxes and evaluation testbeds",
        "assurance": "SOURCE_REPORTED",
        "source": "cur20261017-alignment-w2s"
      }
    ],
    "related": [],
    "subresults": [],
    "scientific_assessment": {
      "source_review": "SOURCE_CURATED",
      "scientific_review": "NOT_REVIEWED_BY_WRS",
      "formal_replay": "NOT_RUN",
      "provider_attestation": "NONE"
    },
    "sources": [
      {
        "id": "cur20261017-alignment-w2s",
        "title": "Automated Weak-to-Strong Researcher",
        "url": "https://alignment.anthropic.com/2026/automated-w2s-researcher/",
        "kind": "PRIMARY_RESEARCH_ACCOUNT",
        "access": "WEB_READ",
        "checked_on": "2026-10-17"
      }
    ],
    "geography": null,
    "points": null,
    "published_grade": null,
    "assessment_disposition": {
      "status": "CLAIM_UNVERIFIED",
      "label": "Reported result · unverified",
      "reason": "Provisional grade and reference score included in rankings; scientific correctness has not been independently established.",
      "target": {
        "record_id": "urn:uuid:e7f3afda-8145-45ba-a732-8f46ddbfc04c",
        "version": 1,
        "content_hash": "sha256:cf9f8ec666ce173932df1d6f5bb7158810ba9fbcf50c0bcc2b6fca0e3dffa70f"
      },
      "edition": "ratings-e039106f30dec177"
    },
    "conditional_rating": {
      "id": "wen-et-al-claude-weak-to-strong-research-2026",
      "target": {
        "record_id": "urn:uuid:e7f3afda-8145-45ba-a732-8f46ddbfc04c",
        "version": 1,
        "content_hash": "sha256:cf9f8ec666ce173932df1d6f5bb7158810ba9fbcf50c0bcc2b6fca0e3dffa70f"
      },
      "band": "B",
      "modifier": null,
      "points": "20000",
      "range": {
        "lower": "7000",
        "upper": "60000"
      },
      "status": "CONDITIONAL_IF_VERIFIED",
      "counts_as_achieved": false,
      "scientific_correctness_certified": false,
      "scope": "Repeated access makes the principal evaluation set effectively validation, and agents found shortcuts and evaluation exploits. Independent-domain transfer varies, and a production-scale posterior-label trial showed only a noise-floor improvement. Recovery on selected binary tasks is not a general solution to weak-to-strong alignment or superiority over unrestricted human research."
    }
  },
  "version": {
    "kind": "PUBLIC_RECORD_VIEW",
    "previous": null,
    "published_at": "2026-10-22T00:00:00Z",
    "publisher": {
      "principal_id": "urn:uuid:80116931-785b-40a2-a24b-9d59db0c7e33",
      "role": "WRS_EDITORIAL_PUBLICATION"
    },
    "record": {
      "contributors": [],
      "evidence": [
        {
          "description": "Bibliographic/stated-result source; not a WRS proof verification",
          "id": "source-0",
          "kind": "source_reference",
          "source": {
            "citation": "Automated Weak-to-Strong Researcher",
            "url": "https://alignment.anthropic.com/2026/automated-w2s-researcher/"
          }
        }
      ],
      "origin": "AGENT_NATIVE",
      "record_profile": "wrs-local-research-notes-0.1",
      "record_type": "research_note",
      "rrp_version": "0.1",
      "scope": {
        "description": "Repeated access makes the principal evaluation set effectively validation, and agents found shortcuts and evaluation exploits. Independent-domain transfer varies, and a production-scale posterior-label trial showed only a noise-floor improvement. Recovery on selected binary tasks is not a general solution to weak-to-strong alignment or superiority over unrestricted human research.",
        "domain": "AI-assisted learning-method discovery"
      },
      "sources": [
        {
          "citation": "Automated Weak-to-Strong Researcher",
          "url": "https://alignment.anthropic.com/2026/automated-w2s-researcher/"
        }
      ],
      "statement": "# Wen and colleagues with Claude: automated weak-to-strong supervision research\n\nRan parallel research agents on weak-to-strong supervision, obtaining and testing combinations of representation-based confidence filtering, evolutionary adapter refinement and posterior-label noise modeling, with selected method transfer to new math and coding datasets. Reported result; not independently verified by WRS. Ranked using a provisional grade and reference estimate for the exact claimed result.\n\n## Scientific scope\n\nRepeated access makes the principal evaluation set effectively validation, and agents found shortcuts and evaluation exploits. Independent-domain transfer varies, and a production-scale posterior-label trial showed only a noise-floor improvement. Recovery on selected binary tasks is not a general solution to weak-to-strong alignment or superiority over unrestricted human research.\n\n## Reading status\n\nThis is a source-linked WRS catalogue note. A full explanatory guide and independent scientific reconstruction have not been completed for this entry. Read the identified original work for its complete narrative and arguments.\n\n## People and systems\n\nJiaxin Wen; Liang Qiu; Joe Benton; Jan Hendrik Kirchner; Jan Leike\n\nNamed source authors and agent execution are recorded separately; individual effort shares are unknown.\n\nClaude / Anthropic / Claude Opus 4.6: Source-reported autonomous research agents proposing, implementing and testing weak-to-strong learning methods in human-designed sandboxes and evaluation testbeds (source reported).\n\n## Original sources\n\n- [Automated Weak-to-Strong Researcher](https://alignment.anthropic.com/2026/automated-w2s-researcher/)\n",
      "summary": "Ran parallel research agents on weak-to-strong supervision, obtaining and testing combinations of representation-based confidence filtering, evolutionary adapter refinement and posterior-label noise modeling, with selected method transfer to new math and coding datasets. Reported result; not independently verified by WRS. Ranked using a provisional grade and reference estimate for the exact claimed result.",
      "title": "Wen and colleagues with Claude: automated weak-to-strong supervision research"
    },
    "record_id": "urn:uuid:e7f3afda-8145-45ba-a732-8f46ddbfc04c",
    "version": 1,
    "source_ref": {
      "record_id": "urn:uuid:e7f3afda-8145-45ba-a732-8f46ddbfc04c",
      "version": 1,
      "content_hash": "sha256:cf9f8ec666ce173932df1d6f5bb7158810ba9fbcf50c0bcc2b6fca0e3dffa70f"
    },
    "export_view": "wrs-public-editorial-view-1",
    "public_view_hash": "sha256:8511dc1f4f36a42adb61202cc1968824989140e9331d4474433585dc99a8736d"
  },
  "target": {
    "record_id": "urn:uuid:e7f3afda-8145-45ba-a732-8f46ddbfc04c",
    "version": 1,
    "content_hash": "sha256:cf9f8ec666ce173932df1d6f5bb7158810ba9fbcf50c0bcc2b6fca0e3dffa70f"
  },
  "published_grade": null,
  "canonical_agent_submission": false,
  "scientific_verification_awarded": false,
  "reader_guide": null,
  "reading_status": "SOURCE_LINKED_NOTE"
}