{
  "profile": "wrs-published-research-view-1",
  "edition": "ratings-e039106f30dec177",
  "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS",
  "entry": {
    "id": "lipman-et-al-flow-matching-generative-modeling-2022",
    "title": "Flow Matching: learn a transport field from simple conditional paths",
    "summary": "Lipman and colleagues train continuous normalizing flows by regressing sample-conditioned vector fields, with the same expected parameter gradients as a marginal flow-matching objective under stated regularity assumptions, avoiding ODE simulation during training.\n\nA generative model can transform random noise into a data-like sample by following a learned motion field. Traditionally, training that field could require costly differential-equation solves. Flow Matching instead constructs simple paths conditioned on individual data samples and learns their velocities by regression. The paper shows when these conditional training targets produce the same expected parameter gradients as the harder-to-compute marginal objective. Gaussian conditional paths include diffusion paths and straight displacement paths, giving designers more control over the training target. The authors report favorable image-quality, likelihood and sampling-cost comparisons in their chosen experiments. Generating samples still requires solving an ODE. Conditional straight paths are optimal between their Gaussian endpoints; the paper explicitly says that the combined marginal field need not be an optimal-transport solution.",
    "boundary": "The method learns a vector field for declared probability paths. Conditional and marginal objectives agree under their stated integrability/regularity conditions; simulation-free training does not remove ODE sampling or optimization costs. Optimal-transport conditional paths are not automatically the globally optimal learned transport. Reported generative metrics are setup-dependent and have not been replayed by WRS.",
    "area": "Generative modeling and probability transport",
    "topic": "Mathematics and computing",
    "year": 2022,
    "lane": "HUMAN",
    "result": "ALGORITHM",
    "package_id": "lipman-et-al-flow-matching-generative-modeling-2022",
    "target": {
      "record_id": "urn:uuid:f2c322bd-9959-4dec-a4e8-a740e25b1fdf",
      "version": 1,
      "content_hash": "sha256:b10d683abf836d76a690ced76baec1c99e7d9806c1b2671ba94d0f60c0a3daac"
    },
    "humans": [
      "Yaron Lipman",
      "Ricky T. Q. Chen",
      "Heli Ben-Hamu",
      "Maximilian Nickel",
      "Matt Le"
    ],
    "authors_complete": true,
    "human_role_note": "Source-listed contributors to this scoped work. Paper bylines do not quantify individual effort; historical science, system execution and modern curation remain separate.",
    "agents": [],
    "related": [
      "denoising-diffusion",
      "shao-et-al-deepseekmath-grpo-2024"
    ],
    "subresults": [],
    "scientific_assessment": {
      "source_review": "SOURCE_CURATED",
      "scientific_review": "NOT_REVIEWED_BY_WRS",
      "formal_replay": "NOT_RUN",
      "provider_attestation": "NONE"
    },
    "sources": [
      {
        "id": "wrs24-flow-matching2023-v2-original",
        "title": "Lipman and colleagues, Flow Matching for Generative Modeling; exact arXiv2210.02747v2",
        "url": "https://arxiv.org/pdf/2210.02747v2"
      },
      {
        "id": "wrs24-flow-matching2023-arxiv-identity",
        "title": "Official arXiv title, complete byline and submission history for Flow Matching v2",
        "url": "https://arxiv.org/abs/2210.02747v2"
      }
    ],
    "points": "80000",
    "published_grade": {
      "id": "lipman-et-al-flow-matching-generative-modeling-2022",
      "entry_id": "lipman-et-al-flow-matching-generative-modeling-2022",
      "band": "A",
      "modifier": null,
      "target": {
        "record_id": "urn:uuid:f2c322bd-9959-4dec-a4e8-a740e25b1fdf",
        "version": 1,
        "content_hash": "sha256:b10d683abf836d76a690ced76baec1c99e7d9806c1b2671ba94d0f60c0a3daac"
      },
      "what_changed": "Conditional vector-field regression yields the marginal FM parameter gradient under exchange/integrability assumptions; Gaussian path design includes conditional straight displacement paths and simulation-free CNF training.",
      "scope": "Conditional vector-field regression yields the marginal FM parameter gradient under exchange/integrability assumptions; Gaussian path design includes conditional straight displacement paths and simulation-free CNF training. The method learns a vector field for declared probability paths. Conditional and marginal objectives agree under their stated integrability/regularity conditions; simulation-free training does not remove ODE sampling or optimization costs. Optimal-transport conditional paths are not automatically the globally optimal learned transport. Reported generative metrics are setup-dependent and have not been replayed by WRS.",
      "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:f2c322bd-9959-4dec-a4e8-a740e25b1fdf",
        "version": 1,
        "content_hash": "sha256:b10d683abf836d76a690ced76baec1c99e7d9806c1b2671ba94d0f60c0a3daac"
      },
      "edition": "ratings-e039106f30dec177"
    }
  },
  "version": {
    "kind": "PUBLIC_RECORD_VIEW",
    "previous": null,
    "published_at": "2026-10-03T00:00:00Z",
    "publisher": {
      "principal_id": "urn:uuid:80116931-785b-40a2-a24b-9d59db0c7e33",
      "role": "WRS_EDITORIAL_PUBLICATION"
    },
    "record": {
      "contributors": [
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Yaron Lipman",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Lipman and colleagues, Flow Matching for Generative Modeling; exact arXiv2210.02747v2",
            "url": "https://arxiv.org/pdf/2210.02747v2"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Ricky T. Q. Chen",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Lipman and colleagues, Flow Matching for Generative Modeling; exact arXiv2210.02747v2",
            "url": "https://arxiv.org/pdf/2210.02747v2"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Heli Ben-Hamu",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Lipman and colleagues, Flow Matching for Generative Modeling; exact arXiv2210.02747v2",
            "url": "https://arxiv.org/pdf/2210.02747v2"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Maximilian Nickel",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Lipman and colleagues, Flow Matching for Generative Modeling; exact arXiv2210.02747v2",
            "url": "https://arxiv.org/pdf/2210.02747v2"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Matt Le",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Lipman and colleagues, Flow Matching for Generative Modeling; exact arXiv2210.02747v2",
            "url": "https://arxiv.org/pdf/2210.02747v2"
          }
        }
      ],
      "evidence": [
        {
          "description": "Bibliographic/stated-result source; not a WRS proof verification",
          "id": "source-0",
          "kind": "source_reference",
          "source": {
            "citation": "Lipman and colleagues, Flow Matching for Generative Modeling; exact arXiv2210.02747v2",
            "url": "https://arxiv.org/pdf/2210.02747v2"
          }
        },
        {
          "description": "Bibliographic/stated-result source; not a WRS proof verification",
          "id": "source-1",
          "kind": "source_reference",
          "source": {
            "citation": "Official arXiv title, complete byline and submission history for Flow Matching v2",
            "url": "https://arxiv.org/abs/2210.02747v2"
          }
        }
      ],
      "origin": "HUMAN_LEGACY",
      "record_profile": "wrs-local-research-notes-0.1",
      "record_type": "research_note",
      "rrp_version": "0.1",
      "scope": {
        "description": "The method learns a vector field for declared probability paths. Conditional and marginal objectives agree under their stated integrability/regularity conditions; simulation-free training does not remove ODE sampling or optimization costs. Optimal-transport conditional paths are not automatically the globally optimal learned transport. Reported generative metrics are setup-dependent and have not been replayed by WRS.",
        "domain": "Generative modeling and probability transport"
      },
      "sources": [
        {
          "citation": "Lipman and colleagues, Flow Matching for Generative Modeling; exact arXiv2210.02747v2",
          "url": "https://arxiv.org/pdf/2210.02747v2"
        },
        {
          "citation": "Official arXiv title, complete byline and submission history for Flow Matching v2",
          "url": "https://arxiv.org/abs/2210.02747v2"
        }
      ],
      "statement": "# Flow Matching: learn a transport field from simple conditional paths\n\nLipman and colleagues train continuous normalizing flows by regressing sample-conditioned vector fields, with the same expected parameter gradients as a marginal flow-matching objective under stated regularity assumptions, avoiding ODE simulation during training.\n\nA generative model can transform random noise into a data-like sample by following a learned motion field. Traditionally, training that field could require costly differential-equation solves. Flow Matching instead constructs simple paths conditioned on individual data samples and learns their velocities by regression. The paper shows when these conditional training targets produce the same expected parameter gradients as the harder-to-compute marginal objective. Gaussian conditional paths include diffusion paths and straight displacement paths, giving designers more control over the training target. The authors report favorable image-quality, likelihood and sampling-cost comparisons in their chosen experiments. Generating samples still requires solving an ODE. Conditional straight paths are optimal between their Gaussian endpoints; the paper explicitly says that the combined marginal field need not be an optimal-transport solution.\n\n## Scientific scope\n\nThe method learns a vector field for declared probability paths. Conditional and marginal objectives agree under their stated integrability/regularity conditions; simulation-free training does not remove ODE sampling or optimization costs. Optimal-transport conditional paths are not automatically the globally optimal learned transport. Reported generative metrics are setup-dependent and have not been replayed by WRS.\n\n## People and systems\n\nYaron Lipman; Ricky T. Q. Chen; Heli Ben-Hamu; Maximilian Nickel; Matt Le\n\n\n\n## Original sources\n\n- [Lipman and colleagues, Flow Matching for Generative Modeling; exact arXiv2210.02747v2](https://arxiv.org/pdf/2210.02747v2)\n- [Official arXiv title, complete byline and submission history for Flow Matching v2](https://arxiv.org/abs/2210.02747v2)\n",
      "summary": "Lipman and colleagues train continuous normalizing flows by regressing sample-conditioned vector fields, with the same expected parameter gradients as a marginal flow-matching objective under stated regularity assumptions, avoiding ODE simulation during training.\n\nA generative model can transform random noise into a data-like sample by following a learned motion field. Traditionally, training that field could require costly differential-equation solves. Flow Matching instead constructs simple paths conditioned on individual data samples and learns their velocities by regression. The paper shows when these conditional training targets produce the same expected parameter gradients as the harder-to-compute marginal objective. Gaussian conditional paths include diffusion paths and straight displacement paths, giving designers more control over the training target. The authors report favorable image-quality, likelihood and sampling-cost comparisons in their chosen experiments. Generating samples still requires solving an ODE. Conditional straight paths are optimal between their Gaussian endpoints; the paper explicitly says that the combined marginal field need not be an optimal-transport solution.",
      "title": "Flow Matching: learn a transport field from simple conditional paths"
    },
    "record_id": "urn:uuid:f2c322bd-9959-4dec-a4e8-a740e25b1fdf",
    "version": 1,
    "source_ref": {
      "record_id": "urn:uuid:f2c322bd-9959-4dec-a4e8-a740e25b1fdf",
      "version": 1,
      "content_hash": "sha256:b10d683abf836d76a690ced76baec1c99e7d9806c1b2671ba94d0f60c0a3daac"
    },
    "export_view": "wrs-public-editorial-view-1",
    "public_view_hash": "sha256:152f8fb68d48926dd3c999cf2f08e2520af3cba7bbdbae853da440b79cc0af68"
  },
  "target": {
    "record_id": "urn:uuid:f2c322bd-9959-4dec-a4e8-a740e25b1fdf",
    "version": 1,
    "content_hash": "sha256:b10d683abf836d76a690ced76baec1c99e7d9806c1b2671ba94d0f60c0a3daac"
  },
  "published_grade": {
    "profile": "wrs-published-grade-1",
    "edition": "2026-10-28.1",
    "id": "lipman-et-al-flow-matching-generative-modeling-2022",
    "entry_id": "lipman-et-al-flow-matching-generative-modeling-2022",
    "revision": 1,
    "title": "Flow Matching: learn a transport field from simple conditional paths",
    "context": "Generative modeling and probability transport",
    "track": "Generative modeling and probability transport",
    "band": "A",
    "modifier": null,
    "status": "PROVISIONAL_EDITORIAL",
    "assessed_on": "2026-10-04",
    "evidence_cutoff": "2026-10-04",
    "achievement_state": "SUPPORTED_SCOPED_RESULT",
    "what_changed": "Conditional vector-field regression yields the marginal FM parameter gradient under exchange/integrability assumptions; Gaussian path design includes conditional straight displacement paths and simulation-free CNF training.",
    "scope": "Conditional vector-field regression yields the marginal FM parameter gradient under exchange/integrability assumptions; Gaussian path design includes conditional straight displacement paths and simulation-free CNF training. The method learns a vector field for declared probability paths. Conditional and marginal objectives agree under their stated integrability/regularity conditions; simulation-free training does not remove ODE sampling or optimization costs. Optimal-transport conditional paths are not automatically the globally optimal learned transport. Reported generative metrics are setup-dependent and have not been replayed by WRS.",
    "evidence_summary": "Source-linked, scoped account of the original work. The published scope states the scientific assumptions and remaining verification limits. WRS has not independently certified the full proof, experimental result or system benchmark.",
    "references": [
      {
        "title": "Lipman and colleagues, Flow Matching for Generative Modeling; exact ar Xiv 2210.02747 v 2",
        "url": "https://arxiv.org/pdf/2210.02747v2",
        "locator": "Original linked work. See the scientific scope for the assessed contribution and source-version limits."
      },
      {
        "title": "Official ar Xiv title, complete byline and submission history for Flow Matching v 2",
        "url": "https://arxiv.org/abs/2210.02747v2",
        "locator": "Original linked work. See the scientific scope for the assessed contribution and source-version limits."
      }
    ],
    "roles": [
      {
        "who": "Yaron Lipman",
        "did": "Original source-listed scientific author of this scoped work; byline does not determine effort fractions.",
        "credit": "SCIENCE"
      },
      {
        "who": "Ricky T. Q. Chen",
        "did": "Original source-listed scientific author of this scoped work; byline does not determine effort fractions.",
        "credit": "SCIENCE"
      },
      {
        "who": "Heli Ben-Hamu",
        "did": "Original source-listed scientific author of this scoped work; byline does not determine effort fractions.",
        "credit": "SCIENCE"
      },
      {
        "who": "Maximilian Nickel",
        "did": "Original source-listed scientific author of this scoped work; byline does not determine effort fractions.",
        "credit": "SCIENCE"
      },
      {
        "who": "Matt Le",
        "did": "Original source-listed scientific author of this scoped work; byline does not determine effort fractions.",
        "credit": "SCIENCE"
      }
    ],
    "target": {
      "record_id": "urn:uuid:f2c322bd-9959-4dec-a4e8-a740e25b1fdf",
      "version": 1,
      "content_hash": "sha256:b10d683abf836d76a690ced76baec1c99e7d9806c1b2671ba94d0f60c0a3daac"
    },
    "scientific_correctness_certified": false,
    "independent_expert_calibration": false,
    "method": "PROPRIETARY",
    "points": "80000"
  },
  "canonical_agent_submission": false,
  "scientific_verification_awarded": false,
  "reader_guide": null,
  "reading_status": "SOURCE_LINKED_NOTE"
}