{
  "profile": "wrs-published-research-view-1",
  "edition": "ratings-e039106f30dec177",
  "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS",
  "entry": {
    "id": "lu-et-al-discopop-loss-discovery-2024",
    "title": "Lu et al.: feedback-driven discovery of preference-optimization objectives",
    "summary": "Iteratively proposed and evaluated preference-training loss code with an LLM, yielding a logistic/exponential blending objective and testing transfer to held-out tasks.",
    "boundary": "DPO, preference learning and programmatic search are inherited. Acquired contribution is objective-discovery framework plus scoped DiscoPOP loss, not evidence that human preferences or safety are universally optimized. Numerical results and generated-code novelty were not independently audited.",
    "area": "Automated optimization-objective design",
    "topic": "Artificial intelligence",
    "year": 2024,
    "lane": "AI",
    "result": "ALGORITHM",
    "package_id": "lu-et-al-discopop-loss-discovery-2024",
    "target": {
      "record_id": "urn:uuid:37c2cc75-45b1-4cfd-a53a-3843ac622ad5",
      "version": 1,
      "content_hash": "sha256:fbd5a30f766b30616940446ea0513db7dd5418e1e21cb656f5cdac970ab234c5"
    },
    "humans": [
      "Chris Lu",
      "Samuel Holt",
      "Claudio Fanconi",
      "Alex J. Chan",
      "Jakob Foerster",
      "Mihaela van der Schaar",
      "Robert Tjarko Lange"
    ],
    "authors_complete": true,
    "human_role_note": "Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.",
    "agents": [
      {
        "provider": "OpenAI",
        "product": "GPT-4",
        "model": "GPT-4",
        "role": "Generated candidate preference-loss code inside a human-designed search, training and evaluation loop; exact service snapshot is not specified by this record.",
        "assurance": "SOURCE_REPORTED",
        "source": "cur20261015-arxiv-2406-08414v1"
      }
    ],
    "related": [],
    "subresults": [],
    "scientific_assessment": {
      "source_review": "SOURCE_CURATED",
      "scientific_review": "NOT_REVIEWED_BY_WRS",
      "formal_replay": "NOT_RUN",
      "provider_attestation": "NONE"
    },
    "sources": [
      {
        "id": "cur20261015-arxiv-2406-08414v1",
        "title": "Discovering Preference Optimization Algorithms with and for Large Language Models",
        "url": "https://arxiv.org/abs/2406.08414v1"
      },
      {
        "id": "cur20261015-support-lu-et-al-discopop-loss-discovery-2024",
        "title": "Lu et al.: feedback-driven discovery of preference-optimization objectives — author/institutional project",
        "url": "https://sakana.ai/llm-squared/"
      }
    ],
    "points": "65000",
    "published_grade": {
      "id": "lu-et-al-discopop-loss-discovery-2024",
      "entry_id": "lu-et-al-discopop-loss-discovery-2024",
      "band": "B",
      "modifier": "+",
      "target": {
        "record_id": "urn:uuid:37c2cc75-45b1-4cfd-a53a-3843ac622ad5",
        "version": 1,
        "content_hash": "sha256:fbd5a30f766b30616940446ea0513db7dd5418e1e21cb656f5cdac970ab234c5"
      },
      "what_changed": "Iteratively proposed and evaluated preference-training loss code with an LLM, yielding a logistic/exponential blending objective and testing transfer to held-out tasks.",
      "scope": "Iteratively proposed and evaluated preference-training loss code with an LLM, yielding a logistic/exponential blending objective and testing transfer to held-out tasks. DPO, preference learning and programmatic search are inherited. Acquired contribution is objective-discovery framework plus scoped DiscoPOP loss, not evidence that human preferences or safety are universally optimized. Numerical results and generated-code novelty were not independently audited.",
      "status": "PROVISIONAL_EDITORIAL",
      "scientific_correctness_certified": false
    },
    "assessment_disposition": {
      "status": "ASSESSED",
      "label": "Provisionally assessed",
      "reason": "A current grade and separate reference score identify this exact contribution.",
      "target": {
        "record_id": "urn:uuid:37c2cc75-45b1-4cfd-a53a-3843ac622ad5",
        "version": 1,
        "content_hash": "sha256:fbd5a30f766b30616940446ea0513db7dd5418e1e21cb656f5cdac970ab234c5"
      },
      "edition": "ratings-e039106f30dec177"
    }
  },
  "version": {
    "kind": "PUBLIC_RECORD_VIEW",
    "previous": null,
    "published_at": "2026-10-16T00:00:00Z",
    "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": "Discovering Preference Optimization Algorithms with and for Large Language Models",
            "url": "https://arxiv.org/abs/2406.08414v1"
          }
        },
        {
          "description": "Bibliographic/stated-result source; not a WRS proof verification",
          "id": "source-1",
          "kind": "source_reference",
          "source": {
            "citation": "Lu et al.: feedback-driven discovery of preference-optimization objectives — author/institutional project",
            "url": "https://sakana.ai/llm-squared/"
          }
        }
      ],
      "origin": "AGENT_NATIVE",
      "record_profile": "wrs-local-research-notes-0.1",
      "record_type": "research_note",
      "rrp_version": "0.1",
      "scope": {
        "description": "DPO, preference learning and programmatic search are inherited. Acquired contribution is objective-discovery framework plus scoped DiscoPOP loss, not evidence that human preferences or safety are universally optimized. Numerical results and generated-code novelty were not independently audited.",
        "domain": "Automated optimization-objective design"
      },
      "sources": [
        {
          "citation": "Discovering Preference Optimization Algorithms with and for Large Language Models",
          "url": "https://arxiv.org/abs/2406.08414v1"
        },
        {
          "citation": "Lu et al.: feedback-driven discovery of preference-optimization objectives — author/institutional project",
          "url": "https://sakana.ai/llm-squared/"
        }
      ],
      "statement": "# Lu et al.: feedback-driven discovery of preference-optimization objectives\n\nIteratively proposed and evaluated preference-training loss code with an LLM, yielding a logistic/exponential blending objective and testing transfer to held-out tasks.\n\n## Scientific scope\n\nDPO, preference learning and programmatic search are inherited. Acquired contribution is objective-discovery framework plus scoped DiscoPOP loss, not evidence that human preferences or safety are universally optimized. Numerical results and generated-code novelty were not independently audited.\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\nChris Lu; Samuel Holt; Claudio Fanconi; Alex J. Chan; Jakob Foerster; Mihaela van der Schaar; Robert Tjarko Lange\n\nSource-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.\n\nGPT-4 / OpenAI / GPT-4: Generated candidate preference-loss code inside a human-designed search, training and evaluation loop; exact service snapshot is not specified by this record. (source reported).\n\n## Original sources\n\n- [Discovering Preference Optimization Algorithms with and for Large Language Models](https://arxiv.org/abs/2406.08414v1)\n- [Lu et al.: feedback-driven discovery of preference-optimization objectives — author/institutional project](https://sakana.ai/llm-squared/)\n",
      "summary": "Iteratively proposed and evaluated preference-training loss code with an LLM, yielding a logistic/exponential blending objective and testing transfer to held-out tasks.",
      "title": "Lu et al.: feedback-driven discovery of preference-optimization objectives"
    },
    "record_id": "urn:uuid:37c2cc75-45b1-4cfd-a53a-3843ac622ad5",
    "version": 1,
    "source_ref": {
      "record_id": "urn:uuid:37c2cc75-45b1-4cfd-a53a-3843ac622ad5",
      "version": 1,
      "content_hash": "sha256:fbd5a30f766b30616940446ea0513db7dd5418e1e21cb656f5cdac970ab234c5"
    },
    "export_view": "wrs-public-editorial-view-1",
    "public_view_hash": "sha256:327a02f744c20170faf1842ea5d136ca7de87f9bd36b60af4f2e8c4610266739"
  },
  "target": {
    "record_id": "urn:uuid:37c2cc75-45b1-4cfd-a53a-3843ac622ad5",
    "version": 1,
    "content_hash": "sha256:fbd5a30f766b30616940446ea0513db7dd5418e1e21cb656f5cdac970ab234c5"
  },
  "published_grade": {
    "profile": "wrs-published-grade-1",
    "edition": "2026-10-28.1",
    "id": "lu-et-al-discopop-loss-discovery-2024",
    "entry_id": "lu-et-al-discopop-loss-discovery-2024",
    "revision": 1,
    "title": "Lu et al.: feedback-driven discovery of preference-optimization objectives",
    "context": "Automated optimization-objective design",
    "track": "AI-assisted algorithm discovery",
    "band": "B",
    "modifier": "+",
    "status": "PROVISIONAL_EDITORIAL",
    "assessed_on": "2026-10-16",
    "evidence_cutoff": "2026-10-16",
    "achievement_state": "SUPPORTED_SCOPED_RESULT",
    "what_changed": "Iteratively proposed and evaluated preference-training loss code with an LLM, yielding a logistic/exponential blending objective and testing transfer to held-out tasks.",
    "scope": "Iteratively proposed and evaluated preference-training loss code with an LLM, yielding a logistic/exponential blending objective and testing transfer to held-out tasks. DPO, preference learning and programmatic search are inherited. Acquired contribution is objective-discovery framework plus scoped DiscoPOP loss, not evidence that human preferences or safety are universally optimized. Numerical results and generated-code novelty were not independently audited.",
    "evidence_summary": "The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. DPO, preference learning and programmatic search are inherited. Acquired contribution is objective-discovery framework plus scoped DiscoPOP loss, not evidence that human preferences or safety are universally optimized. Numerical results and generated-code novelty were not independently audited.",
    "references": [
      {
        "title": "Discovering Preference Optimization Algorithms with and for Large Language Models",
        "url": "https://arxiv.org/abs/2406.08414v1",
        "locator": "Original contribution and stated assumptions, evaluation scope and limitations; source-linked assessment, not independent execution or proof certification."
      },
      {
        "title": "Lu et al.: feedback-driven discovery of preference-optimization objectives — author/institutional project",
        "url": "https://sakana.ai/llm-squared/",
        "locator": "Original contribution and stated assumptions, evaluation scope and limitations; source-linked assessment, not independent execution or proof certification."
      }
    ],
    "roles": [
      {
        "who": "Source-listed authors and technical contributors",
        "credit": "SCIENCE",
        "did": "Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities."
      },
      {
        "who": "GPT-4 / GPT-4",
        "credit": "EXECUTION",
        "did": "Generated candidate preference-loss code inside a human-designed search, training and evaluation loop; exact service snapshot is not specified by this record."
      }
    ],
    "target": {
      "record_id": "urn:uuid:37c2cc75-45b1-4cfd-a53a-3843ac622ad5",
      "version": 1,
      "content_hash": "sha256:fbd5a30f766b30616940446ea0513db7dd5418e1e21cb656f5cdac970ab234c5"
    },
    "scientific_correctness_certified": false,
    "independent_expert_calibration": false,
    "method": "PROPRIETARY",
    "points": "65000"
  },
  "canonical_agent_submission": false,
  "scientific_verification_awarded": false,
  "reader_guide": null,
  "reading_status": "SOURCE_LINKED_NOTE"
}