{
  "profile": "wrs-published-research-view-1",
  "edition": "ratings-e039106f30dec177",
  "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS",
  "entry": {
    "id": "minimax-m1-cispo-reasoning-2025",
    "title": "MiniMax: CISPO and hybrid-attention reasoning training",
    "summary": "Combined efficient hybrid-attention reasoning with CISPO reinforcement learning, which clips importance-sampling weights, and trained on reasoning and software-environment tasks.",
    "boundary": "Original 40K/80K checkpoints are one training contribution. Reported rental cost covers the stated RL run, not pretraining, research or total development. Relative clipping and compute benefits are empirical, not universal optimization guarantees.",
    "area": "Efficient foundation-model architecture and training",
    "topic": "Artificial intelligence",
    "year": 2025,
    "lane": "HUMAN",
    "result": "EMPIRICAL",
    "package_id": "minimax-m1-cispo-reasoning-2025",
    "target": {
      "record_id": "urn:uuid:5c66b328-60a0-4ab9-a615-9c847cc33702",
      "version": 1,
      "content_hash": "sha256:ace9efceaeb167e789ee558c61edda1a7a8f74d8aea2bb595d0ee0e7159eb456"
    },
    "humans": [],
    "authors_complete": false,
    "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": [],
    "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-2506-13585v1",
        "title": "MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention",
        "url": "https://arxiv.org/abs/2506.13585v1"
      },
      {
        "id": "cur20261015-support-minimax-m1-cispo-reasoning-2025",
        "title": "MiniMax: CISPO and hybrid-attention reasoning training — official company record",
        "url": "https://github.com/MiniMax-AI/MiniMax-M1"
      }
    ],
    "points": "65000",
    "published_grade": {
      "id": "minimax-m1-cispo-reasoning-2025",
      "entry_id": "minimax-m1-cispo-reasoning-2025",
      "band": "B",
      "modifier": "+",
      "target": {
        "record_id": "urn:uuid:5c66b328-60a0-4ab9-a615-9c847cc33702",
        "version": 1,
        "content_hash": "sha256:ace9efceaeb167e789ee558c61edda1a7a8f74d8aea2bb595d0ee0e7159eb456"
      },
      "what_changed": "Combined efficient hybrid-attention reasoning with CISPO reinforcement learning, which clips importance-sampling weights, and trained on reasoning and software-environment tasks.",
      "scope": "Combined efficient hybrid-attention reasoning with CISPO reinforcement learning, which clips importance-sampling weights, and trained on reasoning and software-environment tasks. Original 40K/80K checkpoints are one training contribution. Reported rental cost covers the stated RL run, not pretraining, research or total development. Relative clipping and compute benefits are empirical, not universal optimization guarantees.",
      "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:5c66b328-60a0-4ab9-a615-9c847cc33702",
        "version": 1,
        "content_hash": "sha256:ace9efceaeb167e789ee558c61edda1a7a8f74d8aea2bb595d0ee0e7159eb456"
      },
      "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": [
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "organization",
          "name": "MiniMax",
          "note": "Cited historical collaboration, not a platform organization account; individual attribution remains incomplete.",
          "roles": [
            "discoverer"
          ],
          "source": {
            "citation": "MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention",
            "url": "https://arxiv.org/abs/2506.13585v1"
          }
        }
      ],
      "evidence": [
        {
          "description": "Bibliographic/stated-result source; not a WRS proof verification",
          "id": "source-0",
          "kind": "source_reference",
          "source": {
            "citation": "MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention",
            "url": "https://arxiv.org/abs/2506.13585v1"
          }
        },
        {
          "description": "Bibliographic/stated-result source; not a WRS proof verification",
          "id": "source-1",
          "kind": "source_reference",
          "source": {
            "citation": "MiniMax: CISPO and hybrid-attention reasoning training — official company record",
            "url": "https://github.com/MiniMax-AI/MiniMax-M1"
          }
        }
      ],
      "origin": "HUMAN_LEGACY",
      "record_profile": "wrs-local-research-notes-0.1",
      "record_type": "research_note",
      "rrp_version": "0.1",
      "scope": {
        "description": "Original 40K/80K checkpoints are one training contribution. Reported rental cost covers the stated RL run, not pretraining, research or total development. Relative clipping and compute benefits are empirical, not universal optimization guarantees.",
        "domain": "Efficient foundation-model architecture and training"
      },
      "sources": [
        {
          "citation": "MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention",
          "url": "https://arxiv.org/abs/2506.13585v1"
        },
        {
          "citation": "MiniMax: CISPO and hybrid-attention reasoning training — official company record",
          "url": "https://github.com/MiniMax-AI/MiniMax-M1"
        }
      ],
      "statement": "# MiniMax: CISPO and hybrid-attention reasoning training\n\nCombined efficient hybrid-attention reasoning with CISPO reinforcement learning, which clips importance-sampling weights, and trained on reasoning and software-environment tasks.\n\n## Scientific scope\n\nOriginal 40K/80K checkpoints are one training contribution. Reported rental cost covers the stated RL run, not pretraining, research or total development. Relative clipping and compute benefits are empirical, not universal optimization guarantees.\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\n\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\n\n\n## Original sources\n\n- [MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention](https://arxiv.org/abs/2506.13585v1)\n- [MiniMax: CISPO and hybrid-attention reasoning training — official company record](https://github.com/MiniMax-AI/MiniMax-M1)\n",
      "summary": "Combined efficient hybrid-attention reasoning with CISPO reinforcement learning, which clips importance-sampling weights, and trained on reasoning and software-environment tasks.",
      "title": "MiniMax: CISPO and hybrid-attention reasoning training"
    },
    "record_id": "urn:uuid:5c66b328-60a0-4ab9-a615-9c847cc33702",
    "version": 1,
    "source_ref": {
      "record_id": "urn:uuid:5c66b328-60a0-4ab9-a615-9c847cc33702",
      "version": 1,
      "content_hash": "sha256:ace9efceaeb167e789ee558c61edda1a7a8f74d8aea2bb595d0ee0e7159eb456"
    },
    "export_view": "wrs-public-editorial-view-1",
    "public_view_hash": "sha256:53aae54a4fec63d089b73ad2645dc6cab5507817fc641b48f020438f32ee1461"
  },
  "target": {
    "record_id": "urn:uuid:5c66b328-60a0-4ab9-a615-9c847cc33702",
    "version": 1,
    "content_hash": "sha256:ace9efceaeb167e789ee558c61edda1a7a8f74d8aea2bb595d0ee0e7159eb456"
  },
  "published_grade": {
    "profile": "wrs-published-grade-1",
    "edition": "2026-10-28.1",
    "id": "minimax-m1-cispo-reasoning-2025",
    "entry_id": "minimax-m1-cispo-reasoning-2025",
    "revision": 1,
    "title": "MiniMax: CISPO and hybrid-attention reasoning training",
    "context": "Efficient foundation-model architecture and training",
    "track": "Human-designed research method or resource",
    "band": "B",
    "modifier": "+",
    "status": "PROVISIONAL_EDITORIAL",
    "assessed_on": "2026-10-16",
    "evidence_cutoff": "2026-10-16",
    "achievement_state": "SUPPORTED_SCOPED_RESULT",
    "what_changed": "Combined efficient hybrid-attention reasoning with CISPO reinforcement learning, which clips importance-sampling weights, and trained on reasoning and software-environment tasks.",
    "scope": "Combined efficient hybrid-attention reasoning with CISPO reinforcement learning, which clips importance-sampling weights, and trained on reasoning and software-environment tasks. Original 40K/80K checkpoints are one training contribution. Reported rental cost covers the stated RL run, not pretraining, research or total development. Relative clipping and compute benefits are empirical, not universal optimization guarantees.",
    "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. Original 40K/80K checkpoints are one training contribution. Reported rental cost covers the stated RL run, not pretraining, research or total development. Relative clipping and compute benefits are empirical, not universal optimization guarantees.",
    "references": [
      {
        "title": "MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention",
        "url": "https://arxiv.org/abs/2506.13585v1",
        "locator": "Original contribution and stated assumptions, evaluation scope and limitations; source-linked assessment, not independent execution or proof certification."
      },
      {
        "title": "MiniMax: CISPO and hybrid-attention reasoning training — official company record",
        "url": "https://github.com/MiniMax-AI/MiniMax-M1",
        "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."
      }
    ],
    "target": {
      "record_id": "urn:uuid:5c66b328-60a0-4ab9-a615-9c847cc33702",
      "version": 1,
      "content_hash": "sha256:ace9efceaeb167e789ee558c61edda1a7a8f74d8aea2bb595d0ee0e7159eb456"
    },
    "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"
}